Merge pull request #13646 from pollockjj/pre_master_all_squashed

pyisolate-support release refresh

.ci/windows_intel_base_files/run_intel_gpu.bat

@@ -0,0 +1,2 @@
+.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
+pause

.github/workflows/tag-dispatch-cloud.yml

@@ -0,0 +1,45 @@
+name: Tag Dispatch to Cloud
+
+on:
+  push:
+    tags:
+      - 'v*'
+
+jobs:
+  dispatch-cloud:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Send repository dispatch to cloud
+        env:
+          DISPATCH_TOKEN: ${{ secrets.CLOUD_REPO_DISPATCH_TOKEN }}
+          RELEASE_TAG: ${{ github.ref_name }}
+        run: |
+          set -euo pipefail
+
+          if [ -z "${DISPATCH_TOKEN:-}" ]; then
+            echo "::error::CLOUD_REPO_DISPATCH_TOKEN is required but not set."
+            exit 1
+          fi
+
+          RELEASE_URL="https://github.com/${{ github.repository }}/releases/tag/${RELEASE_TAG}"
+
+          PAYLOAD="$(jq -n \
+            --arg release_tag "$RELEASE_TAG" \
+            --arg release_url "$RELEASE_URL" \
+            '{
+              event_type: "comfyui_tag_pushed",
+              client_payload: {
+                release_tag: $release_tag,
+                release_url: $release_url
+              }
+            }')"
+
+          curl -fsSL \
+            -X POST \
+            -H "Accept: application/vnd.github+json" \
+            -H "Content-Type: application/json" \
+            -H "Authorization: Bearer ${DISPATCH_TOKEN}" \
+            https://api.github.com/repos/Comfy-Org/cloud/dispatches \
+            -d "$PAYLOAD"
+
+          echo "✅ Dispatched ComfyUI tag ${RELEASE_TAG} to Comfy-Org/cloud"

.gitignore

@@ -21,6 +21,6 @@ venv*/
 *.log
 web_custom_versions/
 .DS_Store
-openapi.yaml
 filtered-openapi.yaml
 uv.lock
+.pyisolate_venvs/

QUANTIZATION.md

@@ -139,9 +139,9 @@ Example:
   "_quantization_metadata": {
     "format_version": "1.0",
     "layers": {
-      "model.layers.0.mlp.up_proj": "float8_e4m3fn",
-      "model.layers.0.mlp.down_proj": "float8_e4m3fn",
-      "model.layers.1.mlp.up_proj": "float8_e4m3fn"
+      "model.layers.0.mlp.up_proj": {"format": "float8_e4m3fn"},
+      "model.layers.0.mlp.down_proj": {"format": "float8_e4m3fn"},
+      "model.layers.1.mlp.up_proj": {"format": "float8_e4m3fn"}
     }
   }
 }
@@ -165,4 +165,4 @@ Activation quantization (e.g., for FP8 Tensor Core operations) requires `input_s
 3. **Compute scales**: Derive `input_scale` from collected statistics
 4. **Store in checkpoint**: Save `input_scale` parameters alongside weights
 
-The calibration dataset should be representative of your target use case. For diffusion models, this typically means a diverse set of prompts and generation parameters.
+The calibration dataset should be representative of your target use case. For diffusion models, this typically means a diverse set of prompts and generation parameters.

README.md

@@ -195,7 +195,9 @@ The portable above currently comes with python 3.13 and pytorch cuda 13.0. Updat
 
 #### Alternative Downloads:
 
-[Experimental portable for AMD GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z)
+[Portable for AMD GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z)
+
+[Experimental portable for Intel GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_intel.7z)
 
 [Portable with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu126.7z) (Supports Nvidia 10 series and older GPUs).
 

api_server/routes/internal/internal_routes.py

@@ -67,7 +67,7 @@ class InternalRoutes:
                 (entry for entry in os.scandir(directory) if is_visible_file(entry)),
                 key=lambda entry: -entry.stat().st_mtime
             )
-            return web.json_response([entry.name for entry in sorted_files], status=200)
+            return web.json_response([f"{entry.name} [{directory_type}]" for entry in sorted_files], status=200)
 
 
     def get_app(self):

blueprints/.glsl/Glow_30.frag

@@ -2,7 +2,6 @@
 precision mediump float;
 
 uniform sampler2D u_image0;
-uniform vec2 u_resolution;
 uniform int u_int0;      // Blend mode
 uniform int u_int1;      // Color tint
 uniform float u_float0;  // Intensity
@@ -75,7 +74,7 @@ void main() {
     float t0 = threshold - 0.15;
     float t1 = threshold + 0.15;
     
-    vec2 texelSize = 1.0 / u_resolution;
+    vec2 texelSize = 1.0 / vec2(textureSize(u_image0, 0));
     float radius2 = radius * radius;
     
     float sampleScale = clamp(radius * 0.75, 0.35, 1.0);

blueprints/.glsl/Image_Blur_1.frag

@@ -12,7 +12,6 @@ const int RADIAL_SAMPLES = 12;
 const float RADIAL_STRENGTH = 0.0003;
 
 uniform sampler2D u_image0;
-uniform vec2 u_resolution;
 uniform int u_int0;      // Blur type (BLUR_GAUSSIAN, BLUR_BOX, BLUR_RADIAL)
 uniform float u_float0;  // Blur radius/amount
 uniform int u_pass;      // Pass index (0 = horizontal, 1 = vertical)
@@ -25,7 +24,7 @@ float gaussian(float x, float sigma) {
 }
 
 void main() {
-    vec2 texelSize = 1.0 / u_resolution;
+    vec2 texelSize = 1.0 / vec2(textureSize(u_image0, 0));
     float radius = max(u_float0, 0.0);
 
     // Radial (angular) blur - single pass, doesn't use separable

blueprints/.glsl/Sharpen_23.frag

@@ -2,14 +2,13 @@
 precision highp float;
 
 uniform sampler2D u_image0;
-uniform vec2 u_resolution;
 uniform float u_float0;  // strength [0.0 – 2.0] typical: 0.3–1.0
 
 in vec2 v_texCoord;
 layout(location = 0) out vec4 fragColor0;
 
 void main() {
-    vec2 texel = 1.0 / u_resolution;
+    vec2 texel = 1.0 / vec2(textureSize(u_image0, 0));
     
     // Sample center and neighbors
     vec4 center = texture(u_image0, v_texCoord);

blueprints/.glsl/Unsharp_Mask_26.frag

@@ -2,7 +2,6 @@
 precision highp float;
 
 uniform sampler2D u_image0;
-uniform vec2 u_resolution;
 uniform float u_float0;  // amount    [0.0 - 3.0]  typical: 0.5-1.5
 uniform float u_float1;  // radius    [0.5 - 10.0] blur radius in pixels
 uniform float u_float2;  // threshold [0.0 - 0.1]  min difference to sharpen
@@ -19,7 +18,7 @@ float getLuminance(vec3 color) {
 }
 
 void main() {
-    vec2 texel = 1.0 / u_resolution;
+    vec2 texel = 1.0 / vec2(textureSize(u_image0, 0));
     float radius = max(u_float1, 0.5);
     float amount = u_float0;
     float threshold = u_float2;

blueprints/Image Channels.json

@@ -1 +1,322 @@
-{"revision": 0, "last_node_id": 29, "last_link_id": 0, "nodes": [{"id": 29, "type": "4c9d6ea4-b912-40e5-8766-6793a9758c53", "pos": [1970, -230], "size": [180, 86], "flags": {}, "order": 5, "mode": 0, "inputs": [{"label": "image", "localized_name": "images.image0", "name": "images.image0", "type": "IMAGE", "link": null}], "outputs": [{"label": "R", "localized_name": "IMAGE0", "name": "IMAGE0", "type": "IMAGE", "links": []}, {"label": "G", "localized_name": "IMAGE1", "name": "IMAGE1", "type": "IMAGE", "links": []}, {"label": "B", "localized_name": "IMAGE2", "name": "IMAGE2", "type": "IMAGE", "links": []}, {"label": "A", "localized_name": "IMAGE3", "name": "IMAGE3", "type": "IMAGE", "links": []}], "title": "Image Channels", "properties": {"proxyWidgets": []}, "widgets_values": []}], "links": [], "version": 0.4, "definitions": {"subgraphs": [{"id": "4c9d6ea4-b912-40e5-8766-6793a9758c53", "version": 1, "state": {"lastGroupId": 0, "lastNodeId": 28, "lastLinkId": 39, "lastRerouteId": 0}, "revision": 0, "config": {}, "name": "Image Channels", "inputNode": {"id": -10, "bounding": [1820, -185, 120, 60]}, "outputNode": {"id": -20, "bounding": [2460, -215, 120, 120]}, "inputs": [{"id": "3522932b-2d86-4a1f-a02a-cb29f3a9d7fe", "name": "images.image0", "type": "IMAGE", "linkIds": [39], "localized_name": "images.image0", "label": "image", "pos": [1920, -165]}], "outputs": [{"id": "605cb9c3-b065-4d9b-81d2-3ec331889b2b", "name": "IMAGE0", "type": "IMAGE", "linkIds": [26], "localized_name": "IMAGE0", "label": "R", "pos": [2480, -195]}, {"id": "fb44a77e-0522-43e9-9527-82e7465b3596", "name": "IMAGE1", "type": "IMAGE", "linkIds": [27], "localized_name": "IMAGE1", "label": "G", "pos": [2480, -175]}, {"id": "81460ee6-0131-402a-874f-6bf3001fc4ff", "name": "IMAGE2", "type": "IMAGE", "linkIds": [28], "localized_name": "IMAGE2", "label": "B", "pos": [2480, -155]}, {"id": "ae690246-80d4-4951-b1d9-9306d8a77417", "name": "IMAGE3", "type": "IMAGE", "linkIds": [29], "localized_name": "IMAGE3", "label": "A", "pos": [2480, -135]}], "widgets": [], "nodes": [{"id": 23, "type": "GLSLShader", "pos": [2000, -330], "size": [400, 172], "flags": {}, "order": 0, "mode": 0, "inputs": [{"label": "image", "localized_name": "images.image0", "name": "images.image0", "type": "IMAGE", "link": 39}, {"localized_name": "fragment_shader", "name": "fragment_shader", "type": "STRING", "widget": {"name": "fragment_shader"}, "link": null}, {"localized_name": "size_mode", "name": "size_mode", "type": "COMFY_DYNAMICCOMBO_V3", "widget": {"name": "size_mode"}, "link": null}, {"label": "image1", "localized_name": "images.image1", "name": "images.image1", "shape": 7, "type": "IMAGE", "link": null}], "outputs": [{"label": "R", "localized_name": "IMAGE0", "name": "IMAGE0", "type": "IMAGE", "links": [26]}, {"label": "G", "localized_name": "IMAGE1", "name": "IMAGE1", "type": "IMAGE", "links": [27]}, {"label": "B", "localized_name": "IMAGE2", "name": "IMAGE2", "type": "IMAGE", "links": [28]}, {"label": "A", "localized_name": "IMAGE3", "name": "IMAGE3", "type": "IMAGE", "links": [29]}], "properties": {"Node name for S&R": "GLSLShader"}, "widgets_values": ["#version 300 es\nprecision highp float;\n\nuniform sampler2D u_image0;\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\nlayout(location = 1) out vec4 fragColor1;\nlayout(location = 2) out vec4 fragColor2;\nlayout(location = 3) out vec4 fragColor3;\n\nvoid main() {\n  vec4 color = texture(u_image0, v_texCoord);\n  // Output each channel as grayscale to separate render targets\n  fragColor0 = vec4(vec3(color.r), 1.0);  // Red channel\n  fragColor1 = vec4(vec3(color.g), 1.0);  // Green channel\n  fragColor2 = vec4(vec3(color.b), 1.0);  // Blue channel\n  fragColor3 = vec4(vec3(color.a), 1.0);  // Alpha channel\n}\n", "from_input"]}], "groups": [], "links": [{"id": 39, "origin_id": -10, "origin_slot": 0, "target_id": 23, "target_slot": 0, "type": "IMAGE"}, {"id": 26, "origin_id": 23, "origin_slot": 0, "target_id": -20, "target_slot": 0, "type": "IMAGE"}, {"id": 27, "origin_id": 23, "origin_slot": 1, "target_id": -20, "target_slot": 1, "type": "IMAGE"}, {"id": 28, "origin_id": 23, "origin_slot": 2, "target_id": -20, "target_slot": 2, "type": "IMAGE"}, {"id": 29, "origin_id": 23, "origin_slot": 3, "target_id": -20, "target_slot": 3, "type": "IMAGE"}], "extra": {"workflowRendererVersion": "LG"}, "category": "Image Tools/Color adjust"}]}}
+{
+  "revision": 0,
+  "last_node_id": 29,
+  "last_link_id": 0,
+  "nodes": [
+    {
+      "id": 29,
+      "type": "4c9d6ea4-b912-40e5-8766-6793a9758c53",
+      "pos": [
+        1970,
+        -230
+      ],
+      "size": [
+        180,
+        86
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "image",
+          "localized_name": "images.image0",
+          "name": "images.image0",
+          "type": "IMAGE",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "R",
+          "localized_name": "IMAGE0",
+          "name": "IMAGE0",
+          "type": "IMAGE",
+          "links": []
+        },
+        {
+          "label": "G",
+          "localized_name": "IMAGE1",
+          "name": "IMAGE1",
+          "type": "IMAGE",
+          "links": []
+        },
+        {
+          "label": "B",
+          "localized_name": "IMAGE2",
+          "name": "IMAGE2",
+          "type": "IMAGE",
+          "links": []
+        },
+        {
+          "label": "A",
+          "localized_name": "IMAGE3",
+          "name": "IMAGE3",
+          "type": "IMAGE",
+          "links": []
+        }
+      ],
+      "title": "Image Channels",
+      "properties": {
+        "proxyWidgets": []
+      },
+      "widgets_values": []
+    }
+  ],
+  "links": [],
+  "version": 0.4,
+  "definitions": {
+    "subgraphs": [
+      {
+        "id": "4c9d6ea4-b912-40e5-8766-6793a9758c53",
+        "version": 1,
+        "state": {
+          "lastGroupId": 0,
+          "lastNodeId": 28,
+          "lastLinkId": 39,
+          "lastRerouteId": 0
+        },
+        "revision": 0,
+        "config": {},
+        "name": "Image Channels",
+        "inputNode": {
+          "id": -10,
+          "bounding": [
+            1820,
+            -185,
+            120,
+            60
+          ]
+        },
+        "outputNode": {
+          "id": -20,
+          "bounding": [
+            2460,
+            -215,
+            120,
+            120
+          ]
+        },
+        "inputs": [
+          {
+            "id": "3522932b-2d86-4a1f-a02a-cb29f3a9d7fe",
+            "name": "images.image0",
+            "type": "IMAGE",
+            "linkIds": [
+              39
+            ],
+            "localized_name": "images.image0",
+            "label": "image",
+            "pos": [
+              1920,
+              -165
+            ]
+          }
+        ],
+        "outputs": [
+          {
+            "id": "605cb9c3-b065-4d9b-81d2-3ec331889b2b",
+            "name": "IMAGE0",
+            "type": "IMAGE",
+            "linkIds": [
+              26
+            ],
+            "localized_name": "IMAGE0",
+            "label": "R",
+            "pos": [
+              2480,
+              -195
+            ]
+          },
+          {
+            "id": "fb44a77e-0522-43e9-9527-82e7465b3596",
+            "name": "IMAGE1",
+            "type": "IMAGE",
+            "linkIds": [
+              27
+            ],
+            "localized_name": "IMAGE1",
+            "label": "G",
+            "pos": [
+              2480,
+              -175
+            ]
+          },
+          {
+            "id": "81460ee6-0131-402a-874f-6bf3001fc4ff",
+            "name": "IMAGE2",
+            "type": "IMAGE",
+            "linkIds": [
+              28
+            ],
+            "localized_name": "IMAGE2",
+            "label": "B",
+            "pos": [
+              2480,
+              -155
+            ]
+          },
+          {
+            "id": "ae690246-80d4-4951-b1d9-9306d8a77417",
+            "name": "IMAGE3",
+            "type": "IMAGE",
+            "linkIds": [
+              29
+            ],
+            "localized_name": "IMAGE3",
+            "label": "A",
+            "pos": [
+              2480,
+              -135
+            ]
+          }
+        ],
+        "widgets": [],
+        "nodes": [
+          {
+            "id": 23,
+            "type": "GLSLShader",
+            "pos": [
+              2000,
+              -330
+            ],
+            "size": [
+              400,
+              172
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [
+              {
+                "label": "image",
+                "localized_name": "images.image0",
+                "name": "images.image0",
+                "type": "IMAGE",
+                "link": 39
+              },
+              {
+                "localized_name": "fragment_shader",
+                "name": "fragment_shader",
+                "type": "STRING",
+                "widget": {
+                  "name": "fragment_shader"
+                },
+                "link": null
+              },
+              {
+                "localized_name": "size_mode",
+                "name": "size_mode",
+                "type": "COMFY_DYNAMICCOMBO_V3",
+                "widget": {
+                  "name": "size_mode"
+                },
+                "link": null
+              },
+              {
+                "label": "image1",
+                "localized_name": "images.image1",
+                "name": "images.image1",
+                "shape": 7,
+                "type": "IMAGE",
+                "link": null
+              }
+            ],
+            "outputs": [
+              {
+                "label": "R",
+                "localized_name": "IMAGE0",
+                "name": "IMAGE0",
+                "type": "IMAGE",
+                "links": [
+                  26
+                ]
+              },
+              {
+                "label": "G",
+                "localized_name": "IMAGE1",
+                "name": "IMAGE1",
+                "type": "IMAGE",
+                "links": [
+                  27
+                ]
+              },
+              {
+                "label": "B",
+                "localized_name": "IMAGE2",
+                "name": "IMAGE2",
+                "type": "IMAGE",
+                "links": [
+                  28
+                ]
+              },
+              {
+                "label": "A",
+                "localized_name": "IMAGE3",
+                "name": "IMAGE3",
+                "type": "IMAGE",
+                "links": [
+                  29
+                ]
+              }
+            ],
+            "properties": {
+              "Node name for S&R": "GLSLShader"
+            },
+            "widgets_values": [
+              "#version 300 es\nprecision highp float;\n\nuniform sampler2D u_image0;\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\nlayout(location = 1) out vec4 fragColor1;\nlayout(location = 2) out vec4 fragColor2;\nlayout(location = 3) out vec4 fragColor3;\n\nvoid main() {\n  vec4 color = texture(u_image0, v_texCoord);\n  // Output each channel as grayscale to separate render targets\n  fragColor0 = vec4(vec3(color.r), 1.0);  // Red channel\n  fragColor1 = vec4(vec3(color.g), 1.0);  // Green channel\n  fragColor2 = vec4(vec3(color.b), 1.0);  // Blue channel\n  fragColor3 = vec4(vec3(color.a), 1.0);  // Alpha channel\n}\n",
+              "from_input"
+            ]
+          }
+        ],
+        "groups": [],
+        "links": [
+          {
+            "id": 39,
+            "origin_id": -10,
+            "origin_slot": 0,
+            "target_id": 23,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 26,
+            "origin_id": 23,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 27,
+            "origin_id": 23,
+            "origin_slot": 1,
+            "target_id": -20,
+            "target_slot": 1,
+            "type": "IMAGE"
+          },
+          {
+            "id": 28,
+            "origin_id": 23,
+            "origin_slot": 2,
+            "target_id": -20,
+            "target_slot": 2,
+            "type": "IMAGE"
+          },
+          {
+            "id": 29,
+            "origin_id": 23,
+            "origin_slot": 3,
+            "target_id": -20,
+            "target_slot": 3,
+            "type": "IMAGE"
+          }
+        ],
+        "extra": {
+          "workflowRendererVersion": "LG"
+        },
+        "category": "Image Tools/Color adjust"
+      }
+    ]
+  }
+}

blueprints/Prompt Enhance.json

@@ -1 +1,278 @@
-{"revision": 0, "last_node_id": 15, "last_link_id": 0, "nodes": [{"id": 15, "type": "24d8bbfd-39d4-4774-bff0-3de40cc7a471", "pos": [-1490, 2040], "size": [400, 260], "flags": {}, "order": 0, "mode": 0, "inputs": [{"name": "prompt", "type": "STRING", "widget": {"name": "prompt"}, "link": null}, {"label": "reference images", "name": "images", "type": "IMAGE", "link": null}], "outputs": [{"name": "STRING", "type": "STRING", "links": null}], "title": "Prompt Enhance", "properties": {"proxyWidgets": [["-1", "prompt"]], "cnr_id": "comfy-core", "ver": "0.14.1"}, "widgets_values": [""]}], "links": [], "version": 0.4, "definitions": {"subgraphs": [{"id": "24d8bbfd-39d4-4774-bff0-3de40cc7a471", "version": 1, "state": {"lastGroupId": 0, "lastNodeId": 15, "lastLinkId": 14, "lastRerouteId": 0}, "revision": 0, "config": {}, "name": "Prompt Enhance", "inputNode": {"id": -10, "bounding": [-2170, 2110, 138.876953125, 80]}, "outputNode": {"id": -20, "bounding": [-640, 2110, 120, 60]}, "inputs": [{"id": "aeab7216-00e0-4528-a09b-bba50845c5a6", "name": "prompt", "type": "STRING", "linkIds": [11], "pos": [-2051.123046875, 2130]}, {"id": "7b73fd36-aa31-4771-9066-f6c83879994b", "name": "images", "type": "IMAGE", "linkIds": [14], "label": "reference images", "pos": [-2051.123046875, 2150]}], "outputs": [{"id": "c7b0d930-68a1-48d1-b496-0519e5837064", "name": "STRING", "type": "STRING", "linkIds": [13], "pos": [-620, 2130]}], "widgets": [], "nodes": [{"id": 11, "type": "GeminiNode", "pos": [-1560, 1990], "size": [470, 470], "flags": {}, "order": 0, "mode": 0, "inputs": [{"localized_name": "images", "name": "images", "shape": 7, "type": "IMAGE", "link": 14}, {"localized_name": "audio", "name": "audio", "shape": 7, "type": "AUDIO", "link": null}, {"localized_name": "video", "name": "video", "shape": 7, "type": "VIDEO", "link": null}, {"localized_name": "files", "name": "files", "shape": 7, "type": "GEMINI_INPUT_FILES", "link": null}, {"localized_name": "prompt", "name": "prompt", "type": "STRING", "widget": {"name": "prompt"}, "link": 11}, {"localized_name": "model", "name": "model", "type": "COMBO", "widget": {"name": "model"}, "link": null}, {"localized_name": "seed", "name": "seed", "type": "INT", "widget": {"name": "seed"}, "link": null}, {"localized_name": "system_prompt", "name": "system_prompt", "shape": 7, "type": "STRING", "widget": {"name": "system_prompt"}, "link": null}], "outputs": [{"localized_name": "STRING", "name": "STRING", "type": "STRING", "links": [13]}], "properties": {"cnr_id": "comfy-core", "ver": "0.14.1", "Node name for S&R": "GeminiNode"}, "widgets_values": ["", "gemini-3-pro-preview", 42, "randomize", "You are an expert in prompt writing.\nBased on the input, rewrite the user's input into a detailed prompt.\nincluding camera settings, lighting, composition, and style.\nReturn the prompt only"], "color": "#432", "bgcolor": "#653"}], "groups": [], "links": [{"id": 11, "origin_id": -10, "origin_slot": 0, "target_id": 11, "target_slot": 4, "type": "STRING"}, {"id": 13, "origin_id": 11, "origin_slot": 0, "target_id": -20, "target_slot": 0, "type": "STRING"}, {"id": 14, "origin_id": -10, "origin_slot": 1, "target_id": 11, "target_slot": 0, "type": "IMAGE"}], "extra": {"workflowRendererVersion": "LG"}, "category": "Text generation/Prompt enhance"}]}, "extra": {}}
+{
+  "revision": 0,
+  "last_node_id": 15,
+  "last_link_id": 0,
+  "nodes": [
+    {
+      "id": 15,
+      "type": "24d8bbfd-39d4-4774-bff0-3de40cc7a471",
+      "pos": [
+        -1490,
+        2040
+      ],
+      "size": [
+        400,
+        260
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "prompt",
+          "type": "STRING",
+          "widget": {
+            "name": "prompt"
+          },
+          "link": null
+        },
+        {
+          "label": "reference images",
+          "name": "images",
+          "type": "IMAGE",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "STRING",
+          "type": "STRING",
+          "links": null
+        }
+      ],
+      "title": "Prompt Enhance",
+      "properties": {
+        "proxyWidgets": [
+          [
+            "-1",
+            "prompt"
+          ]
+        ],
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1"
+      },
+      "widgets_values": [
+        ""
+      ]
+    }
+  ],
+  "links": [],
+  "version": 0.4,
+  "definitions": {
+    "subgraphs": [
+      {
+        "id": "24d8bbfd-39d4-4774-bff0-3de40cc7a471",
+        "version": 1,
+        "state": {
+          "lastGroupId": 0,
+          "lastNodeId": 15,
+          "lastLinkId": 14,
+          "lastRerouteId": 0
+        },
+        "revision": 0,
+        "config": {},
+        "name": "Prompt Enhance",
+        "inputNode": {
+          "id": -10,
+          "bounding": [
+            -2170,
+            2110,
+            138.876953125,
+            80
+          ]
+        },
+        "outputNode": {
+          "id": -20,
+          "bounding": [
+            -640,
+            2110,
+            120,
+            60
+          ]
+        },
+        "inputs": [
+          {
+            "id": "aeab7216-00e0-4528-a09b-bba50845c5a6",
+            "name": "prompt",
+            "type": "STRING",
+            "linkIds": [
+              11
+            ],
+            "pos": [
+              -2051.123046875,
+              2130
+            ]
+          },
+          {
+            "id": "7b73fd36-aa31-4771-9066-f6c83879994b",
+            "name": "images",
+            "type": "IMAGE",
+            "linkIds": [
+              14
+            ],
+            "label": "reference images",
+            "pos": [
+              -2051.123046875,
+              2150
+            ]
+          }
+        ],
+        "outputs": [
+          {
+            "id": "c7b0d930-68a1-48d1-b496-0519e5837064",
+            "name": "STRING",
+            "type": "STRING",
+            "linkIds": [
+              13
+            ],
+            "pos": [
+              -620,
+              2130
+            ]
+          }
+        ],
+        "widgets": [],
+        "nodes": [
+          {
+            "id": 11,
+            "type": "GeminiNode",
+            "pos": [
+              -1560,
+              1990
+            ],
+            "size": [
+              470,
+              470
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "shape": 7,
+                "type": "IMAGE",
+                "link": 14
+              },
+              {
+                "localized_name": "audio",
+                "name": "audio",
+                "shape": 7,
+                "type": "AUDIO",
+                "link": null
+              },
+              {
+                "localized_name": "video",
+                "name": "video",
+                "shape": 7,
+                "type": "VIDEO",
+                "link": null
+              },
+              {
+                "localized_name": "files",
+                "name": "files",
+                "shape": 7,
+                "type": "GEMINI_INPUT_FILES",
+                "link": null
+              },
+              {
+                "localized_name": "prompt",
+                "name": "prompt",
+                "type": "STRING",
+                "widget": {
+                  "name": "prompt"
+                },
+                "link": 11
+              },
+              {
+                "localized_name": "model",
+                "name": "model",
+                "type": "COMBO",
+                "widget": {
+                  "name": "model"
+                },
+                "link": null
+              },
+              {
+                "localized_name": "seed",
+                "name": "seed",
+                "type": "INT",
+                "widget": {
+                  "name": "seed"
+                },
+                "link": null
+              },
+              {
+                "localized_name": "system_prompt",
+                "name": "system_prompt",
+                "shape": 7,
+                "type": "STRING",
+                "widget": {
+                  "name": "system_prompt"
+                },
+                "link": null
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "STRING",
+                "name": "STRING",
+                "type": "STRING",
+                "links": [
+                  13
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.14.1",
+              "Node name for S&R": "GeminiNode"
+            },
+            "widgets_values": [
+              "",
+              "gemini-3-pro-preview",
+              42,
+              "randomize",
+              "You are an expert in prompt writing.\nBased on the input, rewrite the user's input into a detailed prompt.\nincluding camera settings, lighting, composition, and style.\nReturn the prompt only"
+            ],
+            "color": "#432",
+            "bgcolor": "#653"
+          }
+        ],
+        "groups": [],
+        "links": [
+          {
+            "id": 11,
+            "origin_id": -10,
+            "origin_slot": 0,
+            "target_id": 11,
+            "target_slot": 4,
+            "type": "STRING"
+          },
+          {
+            "id": 13,
+            "origin_id": 11,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 0,
+            "type": "STRING"
+          },
+          {
+            "id": 14,
+            "origin_id": -10,
+            "origin_slot": 1,
+            "target_id": 11,
+            "target_slot": 0,
+            "type": "IMAGE"
+          }
+        ],
+        "extra": {
+          "workflowRendererVersion": "LG"
+        },
+        "category": "Text generation/Prompt enhance"
+      }
+    ]
+  },
+  "extra": {}
+}

blueprints/Sharpen.json

@@ -1 +1,309 @@
-{"revision": 0, "last_node_id": 25, "last_link_id": 0, "nodes": [{"id": 25, "type": "621ba4e2-22a8-482d-a369-023753198b7b", "pos": [4610, -790], "size": [230, 58], "flags": {}, "order": 4, "mode": 0, "inputs": [{"label": "image", "localized_name": "images.image0", "name": "images.image0", "type": "IMAGE", "link": null}], "outputs": [{"label": "IMAGE", "localized_name": "IMAGE0", "name": "IMAGE0", "type": "IMAGE", "links": []}], "title": "Sharpen", "properties": {"proxyWidgets": [["24", "value"]]}, "widgets_values": []}], "links": [], "version": 0.4, "definitions": {"subgraphs": [{"id": "621ba4e2-22a8-482d-a369-023753198b7b", "version": 1, "state": {"lastGroupId": 0, "lastNodeId": 24, "lastLinkId": 36, "lastRerouteId": 0}, "revision": 0, "config": {}, "name": "Sharpen", "inputNode": {"id": -10, "bounding": [4090, -825, 120, 60]}, "outputNode": {"id": -20, "bounding": [5150, -825, 120, 60]}, "inputs": [{"id": "37011fb7-14b7-4e0e-b1a0-6a02e8da1fd7", "name": "images.image0", "type": "IMAGE", "linkIds": [34], "localized_name": "images.image0", "label": "image", "pos": [4190, -805]}], "outputs": [{"id": "e9182b3f-635c-4cd4-a152-4b4be17ae4b9", "name": "IMAGE0", "type": "IMAGE", "linkIds": [35], "localized_name": "IMAGE0", "label": "IMAGE", "pos": [5170, -805]}], "widgets": [], "nodes": [{"id": 24, "type": "PrimitiveFloat", "pos": [4280, -1240], "size": [270, 58], "flags": {}, "order": 0, "mode": 0, "inputs": [{"label": "strength", "localized_name": "value", "name": "value", "type": "FLOAT", "widget": {"name": "value"}, "link": null}], "outputs": [{"localized_name": "FLOAT", "name": "FLOAT", "type": "FLOAT", "links": [36]}], "properties": {"Node name for S&R": "PrimitiveFloat", "min": 0, "max": 3, "precision": 2, "step": 0.05}, "widgets_values": [0.5]}, {"id": 23, "type": "GLSLShader", "pos": [4570, -1240], "size": [370, 192], "flags": {}, "order": 1, "mode": 0, "inputs": [{"label": "image0", "localized_name": "images.image0", "name": "images.image0", "type": "IMAGE", "link": 34}, {"label": "image1", "localized_name": "images.image1", "name": "images.image1", "shape": 7, "type": "IMAGE", "link": null}, {"label": "u_float0", "localized_name": "floats.u_float0", "name": "floats.u_float0", "shape": 7, "type": "FLOAT", "link": 36}, {"label": "u_float1", "localized_name": "floats.u_float1", "name": "floats.u_float1", "shape": 7, "type": "FLOAT", "link": null}, {"label": "u_int0", "localized_name": "ints.u_int0", "name": "ints.u_int0", "shape": 7, "type": "INT", "link": null}, {"localized_name": "fragment_shader", "name": "fragment_shader", "type": "STRING", "widget": {"name": "fragment_shader"}, "link": null}, {"localized_name": "size_mode", "name": "size_mode", "type": "COMFY_DYNAMICCOMBO_V3", "widget": {"name": "size_mode"}, "link": null}], "outputs": [{"localized_name": "IMAGE0", "name": "IMAGE0", "type": "IMAGE", "links": [35]}, {"localized_name": "IMAGE1", "name": "IMAGE1", "type": "IMAGE", "links": null}, {"localized_name": "IMAGE2", "name": "IMAGE2", "type": "IMAGE", "links": null}, {"localized_name": "IMAGE3", "name": "IMAGE3", "type": "IMAGE", "links": null}], "properties": {"Node name for S&R": "GLSLShader"}, "widgets_values": ["#version 300 es\nprecision highp float;\n\nuniform sampler2D u_image0;\nuniform vec2 u_resolution;\nuniform float u_float0;  // strength [0.0 – 2.0] typical: 0.3–1.0\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\n\nvoid main() {\n    vec2 texel = 1.0 / u_resolution;\n    \n    // Sample center and neighbors\n    vec4 center = texture(u_image0, v_texCoord);\n    vec4 top    = texture(u_image0, v_texCoord + vec2( 0.0, -texel.y));\n    vec4 bottom = texture(u_image0, v_texCoord + vec2( 0.0,  texel.y));\n    vec4 left   = texture(u_image0, v_texCoord + vec2(-texel.x,  0.0));\n    vec4 right  = texture(u_image0, v_texCoord + vec2( texel.x,  0.0));\n    \n    // Edge enhancement (Laplacian)\n    vec4 edges = center * 4.0 - top - bottom - left - right;\n    \n    // Add edges back scaled by strength\n    vec4 sharpened = center + edges * u_float0;\n    \n    fragColor0 = vec4(clamp(sharpened.rgb, 0.0, 1.0), center.a);\n}", "from_input"]}], "groups": [], "links": [{"id": 36, "origin_id": 24, "origin_slot": 0, "target_id": 23, "target_slot": 2, "type": "FLOAT"}, {"id": 34, "origin_id": -10, "origin_slot": 0, "target_id": 23, "target_slot": 0, "type": "IMAGE"}, {"id": 35, "origin_id": 23, "origin_slot": 0, "target_id": -20, "target_slot": 0, "type": "IMAGE"}], "extra": {"workflowRendererVersion": "LG"}, "category": "Image Tools/Sharpen"}]}}
+{
+  "revision": 0,
+  "last_node_id": 25,
+  "last_link_id": 0,
+  "nodes": [
+    {
+      "id": 25,
+      "type": "621ba4e2-22a8-482d-a369-023753198b7b",
+      "pos": [
+        4610,
+        -790
+      ],
+      "size": [
+        230,
+        58
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "image",
+          "localized_name": "images.image0",
+          "name": "images.image0",
+          "type": "IMAGE",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "IMAGE",
+          "localized_name": "IMAGE0",
+          "name": "IMAGE0",
+          "type": "IMAGE",
+          "links": []
+        }
+      ],
+      "title": "Sharpen",
+      "properties": {
+        "proxyWidgets": [
+          [
+            "24",
+            "value"
+          ]
+        ]
+      },
+      "widgets_values": []
+    }
+  ],
+  "links": [],
+  "version": 0.4,
+  "definitions": {
+    "subgraphs": [
+      {
+        "id": "621ba4e2-22a8-482d-a369-023753198b7b",
+        "version": 1,
+        "state": {
+          "lastGroupId": 0,
+          "lastNodeId": 24,
+          "lastLinkId": 36,
+          "lastRerouteId": 0
+        },
+        "revision": 0,
+        "config": {},
+        "name": "Sharpen",
+        "inputNode": {
+          "id": -10,
+          "bounding": [
+            4090,
+            -825,
+            120,
+            60
+          ]
+        },
+        "outputNode": {
+          "id": -20,
+          "bounding": [
+            5150,
+            -825,
+            120,
+            60
+          ]
+        },
+        "inputs": [
+          {
+            "id": "37011fb7-14b7-4e0e-b1a0-6a02e8da1fd7",
+            "name": "images.image0",
+            "type": "IMAGE",
+            "linkIds": [
+              34
+            ],
+            "localized_name": "images.image0",
+            "label": "image",
+            "pos": [
+              4190,
+              -805
+            ]
+          }
+        ],
+        "outputs": [
+          {
+            "id": "e9182b3f-635c-4cd4-a152-4b4be17ae4b9",
+            "name": "IMAGE0",
+            "type": "IMAGE",
+            "linkIds": [
+              35
+            ],
+            "localized_name": "IMAGE0",
+            "label": "IMAGE",
+            "pos": [
+              5170,
+              -805
+            ]
+          }
+        ],
+        "widgets": [],
+        "nodes": [
+          {
+            "id": 24,
+            "type": "PrimitiveFloat",
+            "pos": [
+              4280,
+              -1240
+            ],
+            "size": [
+              270,
+              58
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [
+              {
+                "label": "strength",
+                "localized_name": "value",
+                "name": "value",
+                "type": "FLOAT",
+                "widget": {
+                  "name": "value"
+                },
+                "link": null
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "FLOAT",
+                "name": "FLOAT",
+                "type": "FLOAT",
+                "links": [
+                  36
+                ]
+              }
+            ],
+            "properties": {
+              "Node name for S&R": "PrimitiveFloat",
+              "min": 0,
+              "max": 3,
+              "precision": 2,
+              "step": 0.05
+            },
+            "widgets_values": [
+              0.5
+            ]
+          },
+          {
+            "id": 23,
+            "type": "GLSLShader",
+            "pos": [
+              4570,
+              -1240
+            ],
+            "size": [
+              370,
+              192
+            ],
+            "flags": {},
+            "order": 1,
+            "mode": 0,
+            "inputs": [
+              {
+                "label": "image0",
+                "localized_name": "images.image0",
+                "name": "images.image0",
+                "type": "IMAGE",
+                "link": 34
+              },
+              {
+                "label": "image1",
+                "localized_name": "images.image1",
+                "name": "images.image1",
+                "shape": 7,
+                "type": "IMAGE",
+                "link": null
+              },
+              {
+                "label": "u_float0",
+                "localized_name": "floats.u_float0",
+                "name": "floats.u_float0",
+                "shape": 7,
+                "type": "FLOAT",
+                "link": 36
+              },
+              {
+                "label": "u_float1",
+                "localized_name": "floats.u_float1",
+                "name": "floats.u_float1",
+                "shape": 7,
+                "type": "FLOAT",
+                "link": null
+              },
+              {
+                "label": "u_int0",
+                "localized_name": "ints.u_int0",
+                "name": "ints.u_int0",
+                "shape": 7,
+                "type": "INT",
+                "link": null
+              },
+              {
+                "localized_name": "fragment_shader",
+                "name": "fragment_shader",
+                "type": "STRING",
+                "widget": {
+                  "name": "fragment_shader"
+                },
+                "link": null
+              },
+              {
+                "localized_name": "size_mode",
+                "name": "size_mode",
+                "type": "COMFY_DYNAMICCOMBO_V3",
+                "widget": {
+                  "name": "size_mode"
+                },
+                "link": null
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "IMAGE0",
+                "name": "IMAGE0",
+                "type": "IMAGE",
+                "links": [
+                  35
+                ]
+              },
+              {
+                "localized_name": "IMAGE1",
+                "name": "IMAGE1",
+                "type": "IMAGE",
+                "links": null
+              },
+              {
+                "localized_name": "IMAGE2",
+                "name": "IMAGE2",
+                "type": "IMAGE",
+                "links": null
+              },
+              {
+                "localized_name": "IMAGE3",
+                "name": "IMAGE3",
+                "type": "IMAGE",
+                "links": null
+              }
+            ],
+            "properties": {
+              "Node name for S&R": "GLSLShader"
+            },
+            "widgets_values": [
+              "#version 300 es\nprecision highp float;\n\nuniform sampler2D u_image0;\nuniform float u_float0;  // strength [0.0 – 2.0] typical: 0.3–1.0\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\n\nvoid main() {\n    vec2 texel = 1.0 / vec2(textureSize(u_image0, 0));\n    \n    // Sample center and neighbors\n    vec4 center = texture(u_image0, v_texCoord);\n    vec4 top    = texture(u_image0, v_texCoord + vec2( 0.0, -texel.y));\n    vec4 bottom = texture(u_image0, v_texCoord + vec2( 0.0,  texel.y));\n    vec4 left   = texture(u_image0, v_texCoord + vec2(-texel.x,  0.0));\n    vec4 right  = texture(u_image0, v_texCoord + vec2( texel.x,  0.0));\n    \n    // Edge enhancement (Laplacian)\n    vec4 edges = center * 4.0 - top - bottom - left - right;\n    \n    // Add edges back scaled by strength\n    vec4 sharpened = center + edges * u_float0;\n    \n    fragColor0 = vec4(clamp(sharpened.rgb, 0.0, 1.0), center.a);\n}",
+              "from_input"
+            ]
+          }
+        ],
+        "groups": [],
+        "links": [
+          {
+            "id": 36,
+            "origin_id": 24,
+            "origin_slot": 0,
+            "target_id": 23,
+            "target_slot": 2,
+            "type": "FLOAT"
+          },
+          {
+            "id": 34,
+            "origin_id": -10,
+            "origin_slot": 0,
+            "target_id": 23,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 35,
+            "origin_id": 23,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 0,
+            "type": "IMAGE"
+          }
+        ],
+        "extra": {
+          "workflowRendererVersion": "LG"
+        },
+        "category": "Image Tools/Sharpen"
+      }
+    ]
+  }
+}

blueprints/Video Upscale(GAN x4).json

@@ -1 +1,420 @@
-{"revision": 0, "last_node_id": 13, "last_link_id": 0, "nodes": [{"id": 13, "type": "cf95b747-3e17-46cb-8097-cac60ff9b2e1", "pos": [1120, 330], "size": [240, 58], "flags": {}, "order": 3, "mode": 0, "inputs": [{"localized_name": "video", "name": "video", "type": "VIDEO", "link": null}, {"name": "model_name", "type": "COMBO", "widget": {"name": "model_name"}, "link": null}], "outputs": [{"localized_name": "VIDEO", "name": "VIDEO", "type": "VIDEO", "links": []}], "title": "Video Upscale(GAN x4)", "properties": {"proxyWidgets": [["-1", "model_name"]], "cnr_id": "comfy-core", "ver": "0.14.1"}, "widgets_values": ["RealESRGAN_x4plus.safetensors"]}], "links": [], "version": 0.4, "definitions": {"subgraphs": [{"id": "cf95b747-3e17-46cb-8097-cac60ff9b2e1", "version": 1, "state": {"lastGroupId": 0, "lastNodeId": 13, "lastLinkId": 19, "lastRerouteId": 0}, "revision": 0, "config": {}, "name": "Video Upscale(GAN x4)", "inputNode": {"id": -10, "bounding": [550, 460, 120, 80]}, "outputNode": {"id": -20, "bounding": [1490, 460, 120, 60]}, "inputs": [{"id": "666d633e-93e7-42dc-8d11-2b7b99b0f2a6", "name": "video", "type": "VIDEO", "linkIds": [10], "localized_name": "video", "pos": [650, 480]}, {"id": "2e23a087-caa8-4d65-99e6-662761aa905a", "name": "model_name", "type": "COMBO", "linkIds": [19], "pos": [650, 500]}], "outputs": [{"id": "0c1768ea-3ec2-412f-9af6-8e0fa36dae70", "name": "VIDEO", "type": "VIDEO", "linkIds": [15], "localized_name": "VIDEO", "pos": [1510, 480]}], "widgets": [], "nodes": [{"id": 2, "type": "ImageUpscaleWithModel", "pos": [1110, 450], "size": [320, 46], "flags": {}, "order": 1, "mode": 0, "inputs": [{"localized_name": "upscale_model", "name": "upscale_model", "type": "UPSCALE_MODEL", "link": 1}, {"localized_name": "image", "name": "image", "type": "IMAGE", "link": 14}], "outputs": [{"localized_name": "IMAGE", "name": "IMAGE", "type": "IMAGE", "links": [13]}], "properties": {"cnr_id": "comfy-core", "ver": "0.10.0", "Node name for S&R": "ImageUpscaleWithModel"}}, {"id": 11, "type": "CreateVideo", "pos": [1110, 550], "size": [320, 78], "flags": {}, "order": 3, "mode": 0, "inputs": [{"localized_name": "images", "name": "images", "type": "IMAGE", "link": 13}, {"localized_name": "audio", "name": "audio", "shape": 7, "type": "AUDIO", "link": 16}, {"localized_name": "fps", "name": "fps", "type": "FLOAT", "widget": {"name": "fps"}, "link": 12}], "outputs": [{"localized_name": "VIDEO", "name": "VIDEO", "type": "VIDEO", "links": [15]}], "properties": {"cnr_id": "comfy-core", "ver": "0.10.0", "Node name for S&R": "CreateVideo"}, "widgets_values": [30]}, {"id": 10, "type": "GetVideoComponents", "pos": [1110, 330], "size": [320, 70], "flags": {}, "order": 2, "mode": 0, "inputs": [{"localized_name": "video", "name": "video", "type": "VIDEO", "link": 10}], "outputs": [{"localized_name": "images", "name": "images", "type": "IMAGE", "links": [14]}, {"localized_name": "audio", "name": "audio", "type": "AUDIO", "links": [16]}, {"localized_name": "fps", "name": "fps", "type": "FLOAT", "links": [12]}], "properties": {"cnr_id": "comfy-core", "ver": "0.10.0", "Node name for S&R": "GetVideoComponents"}}, {"id": 1, "type": "UpscaleModelLoader", "pos": [750, 450], "size": [280, 60], "flags": {}, "order": 0, "mode": 0, "inputs": [{"localized_name": "model_name", "name": "model_name", "type": "COMBO", "widget": {"name": "model_name"}, "link": 19}], "outputs": [{"localized_name": "UPSCALE_MODEL", "name": "UPSCALE_MODEL", "type": "UPSCALE_MODEL", "links": [1]}], "properties": {"cnr_id": "comfy-core", "ver": "0.10.0", "Node name for S&R": "UpscaleModelLoader", "models": [{"name": "RealESRGAN_x4plus.safetensors", "url": "https://huggingface.co/Comfy-Org/Real-ESRGAN_repackaged/resolve/main/RealESRGAN_x4plus.safetensors", "directory": "upscale_models"}]}, "widgets_values": ["RealESRGAN_x4plus.safetensors"]}], "groups": [], "links": [{"id": 1, "origin_id": 1, "origin_slot": 0, "target_id": 2, "target_slot": 0, "type": "UPSCALE_MODEL"}, {"id": 14, "origin_id": 10, "origin_slot": 0, "target_id": 2, "target_slot": 1, "type": "IMAGE"}, {"id": 13, "origin_id": 2, "origin_slot": 0, "target_id": 11, "target_slot": 0, "type": "IMAGE"}, {"id": 16, "origin_id": 10, "origin_slot": 1, "target_id": 11, "target_slot": 1, "type": "AUDIO"}, {"id": 12, "origin_id": 10, "origin_slot": 2, "target_id": 11, "target_slot": 2, "type": "FLOAT"}, {"id": 10, "origin_id": -10, "origin_slot": 0, "target_id": 10, "target_slot": 0, "type": "VIDEO"}, {"id": 15, "origin_id": 11, "origin_slot": 0, "target_id": -20, "target_slot": 0, "type": "VIDEO"}, {"id": 19, "origin_id": -10, "origin_slot": 1, "target_id": 1, "target_slot": 0, "type": "COMBO"}], "extra": {"workflowRendererVersion": "LG"}, "category": "Video generation and editing/Enhance video"}]}, "extra": {}}
+{
+  "revision": 0,
+  "last_node_id": 13,
+  "last_link_id": 0,
+  "nodes": [
+    {
+      "id": 13,
+      "type": "cf95b747-3e17-46cb-8097-cac60ff9b2e1",
+      "pos": [
+        1120,
+        330
+      ],
+      "size": [
+        240,
+        58
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "video",
+          "name": "video",
+          "type": "VIDEO",
+          "link": null
+        },
+        {
+          "name": "model_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "model_name"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "VIDEO",
+          "name": "VIDEO",
+          "type": "VIDEO",
+          "links": []
+        }
+      ],
+      "title": "Video Upscale(GAN x4)",
+      "properties": {
+        "proxyWidgets": [
+          [
+            "-1",
+            "model_name"
+          ]
+        ],
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1"
+      },
+      "widgets_values": [
+        "RealESRGAN_x4plus.safetensors"
+      ]
+    }
+  ],
+  "links": [],
+  "version": 0.4,
+  "definitions": {
+    "subgraphs": [
+      {
+        "id": "cf95b747-3e17-46cb-8097-cac60ff9b2e1",
+        "version": 1,
+        "state": {
+          "lastGroupId": 0,
+          "lastNodeId": 13,
+          "lastLinkId": 19,
+          "lastRerouteId": 0
+        },
+        "revision": 0,
+        "config": {},
+        "name": "Video Upscale(GAN x4)",
+        "inputNode": {
+          "id": -10,
+          "bounding": [
+            550,
+            460,
+            120,
+            80
+          ]
+        },
+        "outputNode": {
+          "id": -20,
+          "bounding": [
+            1490,
+            460,
+            120,
+            60
+          ]
+        },
+        "inputs": [
+          {
+            "id": "666d633e-93e7-42dc-8d11-2b7b99b0f2a6",
+            "name": "video",
+            "type": "VIDEO",
+            "linkIds": [
+              10
+            ],
+            "localized_name": "video",
+            "pos": [
+              650,
+              480
+            ]
+          },
+          {
+            "id": "2e23a087-caa8-4d65-99e6-662761aa905a",
+            "name": "model_name",
+            "type": "COMBO",
+            "linkIds": [
+              19
+            ],
+            "pos": [
+              650,
+              500
+            ]
+          }
+        ],
+        "outputs": [
+          {
+            "id": "0c1768ea-3ec2-412f-9af6-8e0fa36dae70",
+            "name": "VIDEO",
+            "type": "VIDEO",
+            "linkIds": [
+              15
+            ],
+            "localized_name": "VIDEO",
+            "pos": [
+              1510,
+              480
+            ]
+          }
+        ],
+        "widgets": [],
+        "nodes": [
+          {
+            "id": 2,
+            "type": "ImageUpscaleWithModel",
+            "pos": [
+              1110,
+              450
+            ],
+            "size": [
+              320,
+              46
+            ],
+            "flags": {},
+            "order": 1,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "upscale_model",
+                "name": "upscale_model",
+                "type": "UPSCALE_MODEL",
+                "link": 1
+              },
+              {
+                "localized_name": "image",
+                "name": "image",
+                "type": "IMAGE",
+                "link": 14
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "IMAGE",
+                "name": "IMAGE",
+                "type": "IMAGE",
+                "links": [
+                  13
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.10.0",
+              "Node name for S&R": "ImageUpscaleWithModel"
+            }
+          },
+          {
+            "id": 11,
+            "type": "CreateVideo",
+            "pos": [
+              1110,
+              550
+            ],
+            "size": [
+              320,
+              78
+            ],
+            "flags": {},
+            "order": 3,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "link": 13
+              },
+              {
+                "localized_name": "audio",
+                "name": "audio",
+                "shape": 7,
+                "type": "AUDIO",
+                "link": 16
+              },
+              {
+                "localized_name": "fps",
+                "name": "fps",
+                "type": "FLOAT",
+                "widget": {
+                  "name": "fps"
+                },
+                "link": 12
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "VIDEO",
+                "name": "VIDEO",
+                "type": "VIDEO",
+                "links": [
+                  15
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.10.0",
+              "Node name for S&R": "CreateVideo"
+            },
+            "widgets_values": [
+              30
+            ]
+          },
+          {
+            "id": 10,
+            "type": "GetVideoComponents",
+            "pos": [
+              1110,
+              330
+            ],
+            "size": [
+              320,
+              70
+            ],
+            "flags": {},
+            "order": 2,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "video",
+                "name": "video",
+                "type": "VIDEO",
+                "link": 10
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "links": [
+                  14
+                ]
+              },
+              {
+                "localized_name": "audio",
+                "name": "audio",
+                "type": "AUDIO",
+                "links": [
+                  16
+                ]
+              },
+              {
+                "localized_name": "fps",
+                "name": "fps",
+                "type": "FLOAT",
+                "links": [
+                  12
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.10.0",
+              "Node name for S&R": "GetVideoComponents"
+            }
+          },
+          {
+            "id": 1,
+            "type": "UpscaleModelLoader",
+            "pos": [
+              750,
+              450
+            ],
+            "size": [
+              280,
+              60
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "model_name",
+                "name": "model_name",
+                "type": "COMBO",
+                "widget": {
+                  "name": "model_name"
+                },
+                "link": 19
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "UPSCALE_MODEL",
+                "name": "UPSCALE_MODEL",
+                "type": "UPSCALE_MODEL",
+                "links": [
+                  1
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.10.0",
+              "Node name for S&R": "UpscaleModelLoader",
+              "models": [
+                {
+                  "name": "RealESRGAN_x4plus.safetensors",
+                  "url": "https://huggingface.co/Comfy-Org/Real-ESRGAN_repackaged/resolve/main/RealESRGAN_x4plus.safetensors",
+                  "directory": "upscale_models"
+                }
+              ]
+            },
+            "widgets_values": [
+              "RealESRGAN_x4plus.safetensors"
+            ]
+          }
+        ],
+        "groups": [],
+        "links": [
+          {
+            "id": 1,
+            "origin_id": 1,
+            "origin_slot": 0,
+            "target_id": 2,
+            "target_slot": 0,
+            "type": "UPSCALE_MODEL"
+          },
+          {
+            "id": 14,
+            "origin_id": 10,
+            "origin_slot": 0,
+            "target_id": 2,
+            "target_slot": 1,
+            "type": "IMAGE"
+          },
+          {
+            "id": 13,
+            "origin_id": 2,
+            "origin_slot": 0,
+            "target_id": 11,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 16,
+            "origin_id": 10,
+            "origin_slot": 1,
+            "target_id": 11,
+            "target_slot": 1,
+            "type": "AUDIO"
+          },
+          {
+            "id": 12,
+            "origin_id": 10,
+            "origin_slot": 2,
+            "target_id": 11,
+            "target_slot": 2,
+            "type": "FLOAT"
+          },
+          {
+            "id": 10,
+            "origin_id": -10,
+            "origin_slot": 0,
+            "target_id": 10,
+            "target_slot": 0,
+            "type": "VIDEO"
+          },
+          {
+            "id": 15,
+            "origin_id": 11,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 0,
+            "type": "VIDEO"
+          },
+          {
+            "id": 19,
+            "origin_id": -10,
+            "origin_slot": 1,
+            "target_id": 1,
+            "target_slot": 0,
+            "type": "COMBO"
+          }
+        ],
+        "extra": {
+          "workflowRendererVersion": "LG"
+        },
+        "category": "Video generation and editing/Enhance video"
+      }
+    ]
+  },
+  "extra": {}
+}

comfy/cli_args.py

@@ -184,6 +184,8 @@ parser.add_argument("--disable-api-nodes", action="store_true", help="Disable lo
 
 parser.add_argument("--multi-user", action="store_true", help="Enables per-user storage.")
 
+parser.add_argument("--use-process-isolation", action="store_true", help="Enable process isolation for custom nodes with pyproject.toml manifests containing a [tool.comfy.isolation] section.")
+
 parser.add_argument("--verbose", default='INFO', const='DEBUG', nargs="?", choices=['DEBUG', 'INFO', 'WARNING', 'ERROR', 'CRITICAL'], help='Set the logging level')
 parser.add_argument("--log-stdout", action="store_true", help="Send normal process output to stdout instead of stderr (default).")
 

comfy/hooks.py

@@ -14,6 +14,9 @@ if TYPE_CHECKING:
 import comfy.lora
 import comfy.model_management
 import comfy.patcher_extension
+from comfy.cli_args import args
+import uuid
+import os
 from node_helpers import conditioning_set_values
 
 # #######################################################################################################
@@ -61,8 +64,37 @@ class EnumHookScope(enum.Enum):
     HookedOnly = "hooked_only"
 
 
+_ISOLATION_HOOKREF_MODE = args.use_process_isolation or os.environ.get("PYISOLATE_CHILD") == "1"
+
+
 class _HookRef:
-    pass
+    def __init__(self):
+        if _ISOLATION_HOOKREF_MODE:
+            self._pyisolate_id = str(uuid.uuid4())
+
+    def _ensure_pyisolate_id(self):
+        pyisolate_id = getattr(self, "_pyisolate_id", None)
+        if pyisolate_id is None:
+            pyisolate_id = str(uuid.uuid4())
+            self._pyisolate_id = pyisolate_id
+        return pyisolate_id
+
+    def __eq__(self, other):
+        if not _ISOLATION_HOOKREF_MODE:
+            return self is other
+        if not isinstance(other, _HookRef):
+            return False
+        return self._ensure_pyisolate_id() == other._ensure_pyisolate_id()
+
+    def __hash__(self):
+        if not _ISOLATION_HOOKREF_MODE:
+            return id(self)
+        return hash(self._ensure_pyisolate_id())
+
+    def __str__(self):
+        if not _ISOLATION_HOOKREF_MODE:
+            return super().__str__()
+        return f"PYISOLATE_HOOKREF:{self._ensure_pyisolate_id()}"
 
 
 def default_should_register(hook: Hook, model: ModelPatcher, model_options: dict, target_dict: dict[str], registered: HookGroup):
@@ -168,6 +200,8 @@ class WeightHook(Hook):
                 key_map = comfy.lora.model_lora_keys_clip(model.model, key_map)
             else:
                 key_map = comfy.lora.model_lora_keys_unet(model.model, key_map)
+            if self.weights is None:
+                self.weights = {}
             weights = comfy.lora.load_lora(self.weights, key_map, log_missing=False)
         else:
             if target == EnumWeightTarget.Clip:

comfy/isolation/__init__.py

@@ -0,0 +1,436 @@
+# pylint: disable=consider-using-from-import,cyclic-import,global-statement,global-variable-not-assigned,import-outside-toplevel,logging-fstring-interpolation
+from __future__ import annotations
+import asyncio
+import inspect
+import logging
+import os
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Dict, List, Optional, Set, TYPE_CHECKING
+_IMPORT_TORCH = os.environ.get("PYISOLATE_IMPORT_TORCH", "1") == "1"
+
+load_isolated_node = None
+find_manifest_directories = None
+build_stub_class = None
+get_class_types_for_extension = None
+scan_shm_forensics = None
+start_shm_forensics = None
+
+if _IMPORT_TORCH:
+    import folder_paths
+    from .extension_loader import load_isolated_node
+    from .manifest_loader import find_manifest_directories
+    from .runtime_helpers import build_stub_class, get_class_types_for_extension
+    from .shm_forensics import scan_shm_forensics, start_shm_forensics
+
+if TYPE_CHECKING:
+    from pyisolate import ExtensionManager
+    from .extension_wrapper import ComfyNodeExtension
+
+LOG_PREFIX = "]["
+isolated_node_timings: List[tuple[float, Path, int]] = []
+
+if _IMPORT_TORCH:
+    PYISOLATE_VENV_ROOT = Path(folder_paths.base_path) / ".pyisolate_venvs"
+    PYISOLATE_VENV_ROOT.mkdir(parents=True, exist_ok=True)
+
+logger = logging.getLogger(__name__)
+_WORKFLOW_BOUNDARY_MIN_FREE_VRAM_BYTES = 2 * 1024 * 1024 * 1024
+_MODEL_PATCHER_IDLE_TIMEOUT_MS = 120000
+
+
+def initialize_proxies() -> None:
+    from .child_hooks import is_child_process
+
+    is_child = is_child_process()
+
+    if is_child:
+        from .child_hooks import initialize_child_process
+
+        initialize_child_process()
+    else:
+        from .host_hooks import initialize_host_process
+
+        initialize_host_process()
+        if start_shm_forensics is not None:
+            start_shm_forensics()
+
+
+@dataclass(frozen=True)
+class IsolatedNodeSpec:
+    node_name: str
+    display_name: str
+    stub_class: type
+    module_path: Path
+
+
+_ISOLATED_NODE_SPECS: List[IsolatedNodeSpec] = []
+_CLAIMED_PATHS: Set[Path] = set()
+_ISOLATION_SCAN_ATTEMPTED = False
+_EXTENSION_MANAGERS: List["ExtensionManager"] = []
+_RUNNING_EXTENSIONS: Dict[str, "ComfyNodeExtension"] = {}
+_ISOLATION_BACKGROUND_TASK: Optional["asyncio.Task[List[IsolatedNodeSpec]]"] = None
+_EARLY_START_TIME: Optional[float] = None
+
+
+def start_isolation_loading_early(loop: "asyncio.AbstractEventLoop") -> None:
+    global _ISOLATION_BACKGROUND_TASK, _EARLY_START_TIME
+    if _ISOLATION_BACKGROUND_TASK is not None:
+        return
+    _EARLY_START_TIME = time.perf_counter()
+    _ISOLATION_BACKGROUND_TASK = loop.create_task(initialize_isolation_nodes())
+
+
+async def await_isolation_loading() -> List[IsolatedNodeSpec]:
+    global _ISOLATION_BACKGROUND_TASK, _EARLY_START_TIME
+    if _ISOLATION_BACKGROUND_TASK is not None:
+        specs = await _ISOLATION_BACKGROUND_TASK
+        return specs
+    return await initialize_isolation_nodes()
+
+
+async def initialize_isolation_nodes() -> List[IsolatedNodeSpec]:
+    global _ISOLATED_NODE_SPECS, _ISOLATION_SCAN_ATTEMPTED, _CLAIMED_PATHS
+
+    if _ISOLATED_NODE_SPECS:
+        return _ISOLATED_NODE_SPECS
+
+    if _ISOLATION_SCAN_ATTEMPTED:
+        return []
+
+    _ISOLATION_SCAN_ATTEMPTED = True
+    if find_manifest_directories is None or load_isolated_node is None or build_stub_class is None:
+        return []
+    manifest_entries = find_manifest_directories()
+    _CLAIMED_PATHS = {entry[0].resolve() for entry in manifest_entries}
+
+    if not manifest_entries:
+        return []
+
+    os.environ["PYISOLATE_ISOLATION_ACTIVE"] = "1"
+    concurrency_limit = max(1, (os.cpu_count() or 4) // 2)
+    semaphore = asyncio.Semaphore(concurrency_limit)
+
+    async def load_with_semaphore(
+        node_dir: Path, manifest: Path
+    ) -> List[IsolatedNodeSpec]:
+        async with semaphore:
+            load_start = time.perf_counter()
+            spec_list = await load_isolated_node(
+                node_dir,
+                manifest,
+                logger,
+                lambda name, info, extension: build_stub_class(
+                    name,
+                    info,
+                    extension,
+                    _RUNNING_EXTENSIONS,
+                    logger,
+                ),
+                PYISOLATE_VENV_ROOT,
+                _EXTENSION_MANAGERS,
+            )
+            spec_list = [
+                IsolatedNodeSpec(
+                    node_name=node_name,
+                    display_name=display_name,
+                    stub_class=stub_cls,
+                    module_path=node_dir,
+                )
+                for node_name, display_name, stub_cls in spec_list
+            ]
+            isolated_node_timings.append(
+                (time.perf_counter() - load_start, node_dir, len(spec_list))
+            )
+            return spec_list
+
+    tasks = [
+        load_with_semaphore(node_dir, manifest)
+        for node_dir, manifest in manifest_entries
+    ]
+    results = await asyncio.gather(*tasks, return_exceptions=True)
+
+    specs: List[IsolatedNodeSpec] = []
+    for result in results:
+        if isinstance(result, Exception):
+            logger.error(
+                "%s Isolated node failed during startup; continuing: %s",
+                LOG_PREFIX,
+                result,
+            )
+            continue
+        specs.extend(result)
+
+    _ISOLATED_NODE_SPECS = specs
+    return list(_ISOLATED_NODE_SPECS)
+
+
+def _get_class_types_for_extension(extension_name: str) -> Set[str]:
+    """Get all node class types (node names) belonging to an extension."""
+    extension = _RUNNING_EXTENSIONS.get(extension_name)
+    if not extension:
+        return set()
+
+    ext_path = Path(extension.module_path)
+    class_types = set()
+    for spec in _ISOLATED_NODE_SPECS:
+        if spec.module_path.resolve() == ext_path.resolve():
+            class_types.add(spec.node_name)
+
+    return class_types
+
+
+async def notify_execution_graph(needed_class_types: Set[str], caches: list | None = None) -> None:
+    """Evict running extensions not needed for current execution.
+
+    When *caches* is provided, cache entries for evicted extensions' node
+    class_types are invalidated to prevent stale ``RemoteObjectHandle``
+    references from surviving in the output cache.
+    """
+    await wait_for_model_patcher_quiescence(
+        timeout_ms=_MODEL_PATCHER_IDLE_TIMEOUT_MS,
+        fail_loud=True,
+        marker="ISO:notify_graph_wait_idle",
+    )
+
+    evicted_class_types: Set[str] = set()
+
+    async def _stop_extension(
+        ext_name: str, extension: "ComfyNodeExtension", reason: str
+    ) -> None:
+        # Collect class_types BEFORE stopping so we can invalidate cache entries.
+        ext_class_types = _get_class_types_for_extension(ext_name)
+        evicted_class_types.update(ext_class_types)
+        logger.info("%s ISO:eject_start ext=%s reason=%s", LOG_PREFIX, ext_name, reason)
+        logger.debug("%s ISO:stop_start ext=%s", LOG_PREFIX, ext_name)
+        stop_result = extension.stop()
+        if inspect.isawaitable(stop_result):
+            await stop_result
+        _RUNNING_EXTENSIONS.pop(ext_name, None)
+        logger.debug("%s ISO:stop_done ext=%s", LOG_PREFIX, ext_name)
+        if scan_shm_forensics is not None:
+            scan_shm_forensics("ISO:stop_extension", refresh_model_context=True)
+
+    if scan_shm_forensics is not None:
+        scan_shm_forensics("ISO:notify_graph_start", refresh_model_context=True)
+    isolated_class_types_in_graph = needed_class_types.intersection(
+        {spec.node_name for spec in _ISOLATED_NODE_SPECS}
+    )
+    graph_uses_isolation = bool(isolated_class_types_in_graph)
+    logger.debug(
+        "%s ISO:notify_graph_start running=%d needed=%d",
+        LOG_PREFIX,
+        len(_RUNNING_EXTENSIONS),
+        len(needed_class_types),
+    )
+    if graph_uses_isolation:
+        for ext_name, extension in list(_RUNNING_EXTENSIONS.items()):
+            ext_class_types = _get_class_types_for_extension(ext_name)
+
+            # If NONE of this extension's nodes are in the execution graph -> evict.
+            if not ext_class_types.intersection(needed_class_types):
+                await _stop_extension(
+                    ext_name,
+                    extension,
+                    "isolated custom_node not in execution graph, evicting",
+                )
+    else:
+        logger.debug(
+            "%s ISO:notify_graph_skip_evict running=%d reason=no isolated nodes in graph",
+            LOG_PREFIX,
+            len(_RUNNING_EXTENSIONS),
+        )
+
+    # Isolated child processes add steady VRAM pressure; reclaim host-side models
+    # at workflow boundaries so subsequent host nodes (e.g. CLIP encode) keep headroom.
+    try:
+        import comfy.model_management as model_management
+
+        device = model_management.get_torch_device()
+        if getattr(device, "type", None) == "cuda":
+            required = max(
+                model_management.minimum_inference_memory(),
+                _WORKFLOW_BOUNDARY_MIN_FREE_VRAM_BYTES,
+            )
+            free_before = model_management.get_free_memory(device)
+            if free_before < required and _RUNNING_EXTENSIONS and graph_uses_isolation:
+                for ext_name, extension in list(_RUNNING_EXTENSIONS.items()):
+                    await _stop_extension(
+                        ext_name,
+                        extension,
+                        f"boundary low-vram restart (free={int(free_before)} target={int(required)})",
+                    )
+            if model_management.get_free_memory(device) < required:
+                model_management.unload_all_models()
+            model_management.cleanup_models_gc()
+            model_management.cleanup_models()
+            if model_management.get_free_memory(device) < required:
+                model_management.free_memory(required, device, for_dynamic=False)
+                model_management.soft_empty_cache()
+    except Exception:
+        logger.debug(
+            "%s workflow-boundary host VRAM relief failed", LOG_PREFIX, exc_info=True
+        )
+    finally:
+        # Invalidate cached outputs for evicted extensions so stale
+        # RemoteObjectHandle references are not served from cache.
+        if evicted_class_types and caches:
+            total_invalidated = 0
+            for cache in caches:
+                if hasattr(cache, "invalidate_by_class_types"):
+                    total_invalidated += cache.invalidate_by_class_types(
+                        evicted_class_types
+                    )
+            if total_invalidated > 0:
+                logger.info(
+                    "%s ISO:cache_invalidated count=%d class_types=%s",
+                    LOG_PREFIX,
+                    total_invalidated,
+                    evicted_class_types,
+                )
+        scan_shm_forensics("ISO:notify_graph_done", refresh_model_context=True)
+        logger.debug(
+            "%s ISO:notify_graph_done running=%d", LOG_PREFIX, len(_RUNNING_EXTENSIONS)
+        )
+
+
+async def flush_running_extensions_transport_state() -> int:
+    await wait_for_model_patcher_quiescence(
+        timeout_ms=_MODEL_PATCHER_IDLE_TIMEOUT_MS,
+        fail_loud=True,
+        marker="ISO:flush_transport_wait_idle",
+    )
+    total_flushed = 0
+    for ext_name, extension in list(_RUNNING_EXTENSIONS.items()):
+        flush_fn = getattr(extension, "flush_transport_state", None)
+        if not callable(flush_fn):
+            continue
+        try:
+            flushed = await flush_fn()
+            if isinstance(flushed, int):
+                total_flushed += flushed
+                if flushed > 0:
+                    logger.debug(
+                        "%s %s workflow-end flush released=%d",
+                        LOG_PREFIX,
+                        ext_name,
+                        flushed,
+                    )
+        except Exception:
+            logger.debug(
+                "%s %s workflow-end flush failed", LOG_PREFIX, ext_name, exc_info=True
+            )
+    scan_shm_forensics(
+        "ISO:flush_running_extensions_transport_state", refresh_model_context=True
+    )
+    return total_flushed
+
+
+async def wait_for_model_patcher_quiescence(
+    timeout_ms: int = _MODEL_PATCHER_IDLE_TIMEOUT_MS,
+    *,
+    fail_loud: bool = False,
+    marker: str = "ISO:wait_model_patcher_idle",
+) -> bool:
+    try:
+        from comfy.isolation.model_patcher_proxy_registry import ModelPatcherRegistry
+
+        registry = ModelPatcherRegistry()
+        start = time.perf_counter()
+        idle = await registry.wait_all_idle(timeout_ms)
+        elapsed_ms = (time.perf_counter() - start) * 1000.0
+        if idle:
+            logger.debug(
+                "%s %s idle=1 timeout_ms=%d elapsed_ms=%.3f",
+                LOG_PREFIX,
+                marker,
+                timeout_ms,
+                elapsed_ms,
+            )
+            return True
+
+        states = await registry.get_all_operation_states()
+        logger.error(
+            "%s %s idle_timeout timeout_ms=%d elapsed_ms=%.3f states=%s",
+            LOG_PREFIX,
+            marker,
+            timeout_ms,
+            elapsed_ms,
+            states,
+        )
+        if fail_loud:
+            raise TimeoutError(
+                f"ModelPatcherRegistry did not quiesce within {timeout_ms} ms"
+            )
+        return False
+    except Exception:
+        if fail_loud:
+            raise
+        logger.debug("%s %s failed", LOG_PREFIX, marker, exc_info=True)
+        return False
+
+
+def get_claimed_paths() -> Set[Path]:
+    return _CLAIMED_PATHS
+
+
+def update_rpc_event_loops(loop: "asyncio.AbstractEventLoop | None" = None) -> None:
+    """Update all active RPC instances with the current event loop.
+
+    This MUST be called at the start of each workflow execution to ensure
+    RPC calls are scheduled on the correct event loop. This handles the case
+    where asyncio.run() creates a new event loop for each workflow.
+
+    Args:
+        loop: The event loop to use. If None, uses asyncio.get_running_loop().
+    """
+    if loop is None:
+        try:
+            loop = asyncio.get_running_loop()
+        except RuntimeError:
+            loop = asyncio.get_event_loop()
+
+    update_count = 0
+
+    # Update RPCs from ExtensionManagers
+    for manager in _EXTENSION_MANAGERS:
+        if not hasattr(manager, "extensions"):
+            continue
+        for name, extension in manager.extensions.items():
+            if hasattr(extension, "rpc") and extension.rpc is not None:
+                if hasattr(extension.rpc, "update_event_loop"):
+                    extension.rpc.update_event_loop(loop)
+                    update_count += 1
+                    logger.debug(f"{LOG_PREFIX}Updated loop on extension '{name}'")
+
+    # Also update RPCs from running extensions (they may have direct RPC refs)
+    for name, extension in _RUNNING_EXTENSIONS.items():
+        if hasattr(extension, "rpc") and extension.rpc is not None:
+            if hasattr(extension.rpc, "update_event_loop"):
+                extension.rpc.update_event_loop(loop)
+                update_count += 1
+                logger.debug(f"{LOG_PREFIX}Updated loop on running extension '{name}'")
+
+    if update_count > 0:
+        logger.debug(f"{LOG_PREFIX}Updated event loop on {update_count} RPC instances")
+    else:
+        logger.debug(
+            f"{LOG_PREFIX}No RPC instances found to update (managers={len(_EXTENSION_MANAGERS)}, running={len(_RUNNING_EXTENSIONS)})"
+        )
+
+
+__all__ = [
+    "LOG_PREFIX",
+    "initialize_proxies",
+    "initialize_isolation_nodes",
+    "start_isolation_loading_early",
+    "await_isolation_loading",
+    "notify_execution_graph",
+    "flush_running_extensions_transport_state",
+    "wait_for_model_patcher_quiescence",
+    "get_claimed_paths",
+    "update_rpc_event_loops",
+    "IsolatedNodeSpec",
+    "get_class_types_for_extension",
+]

comfy/isolation/adapter.py

@@ -0,0 +1,864 @@
+# pylint: disable=import-outside-toplevel,logging-fstring-interpolation,protected-access,raise-missing-from,useless-return,wrong-import-position
+from __future__ import annotations
+
+import logging
+import os
+import inspect
+from pathlib import Path
+from typing import Any, Dict, List, Optional, cast
+
+from pyisolate.interfaces import IsolationAdapter, SerializerRegistryProtocol  # type: ignore[import-untyped]
+from pyisolate._internal.rpc_protocol import AsyncRPC, ProxiedSingleton  # type: ignore[import-untyped]
+
+_IMPORT_TORCH = os.environ.get("PYISOLATE_IMPORT_TORCH", "1") == "1"
+
+# Singleton proxies that do NOT transitively import torch/PIL/psutil/aiohttp.
+# Safe to import in sealed workers without host framework modules.
+from comfy.isolation.proxies.folder_paths_proxy import FolderPathsProxy
+from comfy.isolation.proxies.helper_proxies import HelperProxiesService
+from comfy.isolation.proxies.web_directory_proxy import WebDirectoryProxy
+
+# Singleton proxies that transitively import torch, PIL, or heavy host modules.
+# Only available when torch/host framework is present.
+CLIPProxy = None
+CLIPRegistry = None
+ModelPatcherProxy = None
+ModelPatcherRegistry = None
+ModelSamplingProxy = None
+ModelSamplingRegistry = None
+VAEProxy = None
+VAERegistry = None
+FirstStageModelRegistry = None
+ModelManagementProxy = None
+PromptServerService = None
+ProgressProxy = None
+UtilsProxy = None
+_HAS_TORCH_PROXIES = False
+if _IMPORT_TORCH:
+    from comfy.isolation.clip_proxy import CLIPProxy, CLIPRegistry
+    from comfy.isolation.model_patcher_proxy import (
+        ModelPatcherProxy,
+        ModelPatcherRegistry,
+    )
+    from comfy.isolation.model_sampling_proxy import (
+        ModelSamplingProxy,
+        ModelSamplingRegistry,
+    )
+    from comfy.isolation.vae_proxy import VAEProxy, VAERegistry, FirstStageModelRegistry
+    from comfy.isolation.proxies.model_management_proxy import ModelManagementProxy
+    from comfy.isolation.proxies.prompt_server_impl import PromptServerService
+    from comfy.isolation.proxies.progress_proxy import ProgressProxy
+    from comfy.isolation.proxies.utils_proxy import UtilsProxy
+    _HAS_TORCH_PROXIES = True
+
+logger = logging.getLogger(__name__)
+
+# Force /dev/shm for shared memory (bwrap makes /tmp private)
+import tempfile
+
+if os.path.exists("/dev/shm"):
+    # Only override if not already set or if default is not /dev/shm
+    current_tmp = tempfile.gettempdir()
+    if not current_tmp.startswith("/dev/shm"):
+        logger.debug(
+            f"Configuring shared memory: Changing TMPDIR from {current_tmp} to /dev/shm"
+        )
+        os.environ["TMPDIR"] = "/dev/shm"
+        tempfile.tempdir = None  # Clear cache to force re-evaluation
+
+
+class ComfyUIAdapter(IsolationAdapter):
+    # ComfyUI-specific IsolationAdapter implementation
+
+    @property
+    def identifier(self) -> str:
+        return "comfyui"
+
+    def get_path_config(self, module_path: str) -> Optional[Dict[str, Any]]:
+        if "ComfyUI" in module_path and "custom_nodes" in module_path:
+            parts = module_path.split("ComfyUI")
+            if len(parts) > 1:
+                comfy_root = parts[0] + "ComfyUI"
+                return {
+                    "preferred_root": comfy_root,
+                    "additional_paths": [
+                        os.path.join(comfy_root, "custom_nodes"),
+                        os.path.join(comfy_root, "comfy"),
+                    ],
+                    "filtered_subdirs": ["comfy", "app", "comfy_execution", "utils"],
+                }
+        return None
+
+    def get_sandbox_system_paths(self) -> Optional[List[str]]:
+        """Returns required application paths to mount in the sandbox."""
+        # By inspecting where our adapter is loaded from, we can determine the comfy root
+        adapter_file = inspect.getfile(self.__class__)
+        # adapter_file = /home/johnj/ComfyUI/comfy/isolation/adapter.py
+        comfy_root = os.path.dirname(os.path.dirname(os.path.dirname(adapter_file)))
+        if os.path.exists(comfy_root):
+            return [comfy_root]
+        return None
+
+    def setup_child_environment(self, snapshot: Dict[str, Any]) -> None:
+        comfy_root = snapshot.get("preferred_root")
+        if not comfy_root:
+            return
+
+        requirements_path = Path(comfy_root) / "requirements.txt"
+        if requirements_path.exists():
+            import re
+
+            for line in requirements_path.read_text().splitlines():
+                line = line.strip()
+                if not line or line.startswith("#"):
+                    continue
+                pkg_name = re.split(r"[<>=!~\[]", line)[0].strip()
+                if pkg_name:
+                    logging.getLogger(pkg_name).setLevel(logging.ERROR)
+
+    def register_serializers(self, registry: SerializerRegistryProtocol) -> None:
+        if not _IMPORT_TORCH:
+            # Sealed worker without torch — register torch-free TensorValue handler
+            # so IMAGE/MASK/LATENT tensors arrive as numpy arrays, not raw dicts.
+            import numpy as np
+
+            _TORCH_DTYPE_TO_NUMPY = {
+                "torch.float32": np.float32,
+                "torch.float64": np.float64,
+                "torch.float16": np.float16,
+                "torch.bfloat16": np.float32,  # numpy has no bfloat16; upcast
+                "torch.int32": np.int32,
+                "torch.int64": np.int64,
+                "torch.int16": np.int16,
+                "torch.int8": np.int8,
+                "torch.uint8": np.uint8,
+                "torch.bool": np.bool_,
+            }
+
+            def _deserialize_tensor_value(data: Dict[str, Any]) -> Any:
+                dtype_str = data["dtype"]
+                np_dtype = _TORCH_DTYPE_TO_NUMPY.get(dtype_str, np.float32)
+                shape = tuple(data["tensor_size"])
+                arr = np.array(data["data"], dtype=np_dtype).reshape(shape)
+                return arr
+
+            _NUMPY_TO_TORCH_DTYPE = {
+                np.float32: "torch.float32",
+                np.float64: "torch.float64",
+                np.float16: "torch.float16",
+                np.int32: "torch.int32",
+                np.int64: "torch.int64",
+                np.int16: "torch.int16",
+                np.int8: "torch.int8",
+                np.uint8: "torch.uint8",
+                np.bool_: "torch.bool",
+            }
+
+            def _serialize_tensor_value(obj: Any) -> Dict[str, Any]:
+                arr = np.asarray(obj, dtype=np.float32) if obj.dtype not in _NUMPY_TO_TORCH_DTYPE else np.asarray(obj)
+                dtype_str = _NUMPY_TO_TORCH_DTYPE.get(arr.dtype.type, "torch.float32")
+                return {
+                    "__type__": "TensorValue",
+                    "dtype": dtype_str,
+                    "tensor_size": list(arr.shape),
+                    "requires_grad": False,
+                    "data": arr.tolist(),
+                }
+
+            registry.register("TensorValue", _serialize_tensor_value, _deserialize_tensor_value, data_type=True)
+            # ndarray output from sealed workers serializes as TensorValue for host torch reconstruction
+            registry.register("ndarray", _serialize_tensor_value, _deserialize_tensor_value, data_type=True)
+            return
+
+        import torch
+
+        def serialize_device(obj: Any) -> Dict[str, Any]:
+            return {"__type__": "device", "device_str": str(obj)}
+
+        def deserialize_device(data: Dict[str, Any]) -> Any:
+            return torch.device(data["device_str"])
+
+        registry.register("device", serialize_device, deserialize_device)
+
+        _VALID_DTYPES = {
+            "float16", "float32", "float64", "bfloat16",
+            "int8", "int16", "int32", "int64",
+            "uint8", "bool",
+        }
+
+        def serialize_dtype(obj: Any) -> Dict[str, Any]:
+            return {"__type__": "dtype", "dtype_str": str(obj)}
+
+        def deserialize_dtype(data: Dict[str, Any]) -> Any:
+            dtype_name = data["dtype_str"].replace("torch.", "")
+            if dtype_name not in _VALID_DTYPES:
+                raise ValueError(f"Invalid dtype: {data['dtype_str']}")
+            return getattr(torch, dtype_name)
+
+        registry.register("dtype", serialize_dtype, deserialize_dtype)
+
+        from comfy_api.latest._io import FolderType
+        from comfy_api.latest._ui import SavedImages, SavedResult
+
+        def serialize_saved_result(obj: Any) -> Dict[str, Any]:
+            return {
+                "__type__": "SavedResult",
+                "filename": obj.filename,
+                "subfolder": obj.subfolder,
+                "folder_type": obj.type.value,
+            }
+
+        def deserialize_saved_result(data: Dict[str, Any]) -> Any:
+            if isinstance(data, SavedResult):
+                return data
+            folder_type = data["folder_type"] if "folder_type" in data else data["type"]
+            return SavedResult(
+                filename=data["filename"],
+                subfolder=data["subfolder"],
+                type=FolderType(folder_type),
+            )
+
+        registry.register(
+            "SavedResult",
+            serialize_saved_result,
+            deserialize_saved_result,
+            data_type=True,
+        )
+
+        def serialize_saved_images(obj: Any) -> Dict[str, Any]:
+            return {
+                "__type__": "SavedImages",
+                "results": [serialize_saved_result(result) for result in obj.results],
+                "is_animated": obj.is_animated,
+            }
+
+        def deserialize_saved_images(data: Dict[str, Any]) -> Any:
+            return SavedImages(
+                results=[deserialize_saved_result(result) for result in data["results"]],
+                is_animated=data.get("is_animated", False),
+            )
+
+        registry.register(
+            "SavedImages",
+            serialize_saved_images,
+            deserialize_saved_images,
+            data_type=True,
+        )
+
+        def serialize_model_patcher(obj: Any) -> Dict[str, Any]:
+            # Child-side: must already have _instance_id (proxy)
+            if os.environ.get("PYISOLATE_CHILD") == "1":
+                if hasattr(obj, "_instance_id"):
+                    return {"__type__": "ModelPatcherRef", "model_id": obj._instance_id}
+                raise RuntimeError(
+                    f"ModelPatcher in child lacks _instance_id: "
+                    f"{type(obj).__module__}.{type(obj).__name__}"
+                )
+            # Host-side: register with registry
+            if hasattr(obj, "_instance_id"):
+                return {"__type__": "ModelPatcherRef", "model_id": obj._instance_id}
+            model_id = ModelPatcherRegistry().register(obj)
+            return {"__type__": "ModelPatcherRef", "model_id": model_id}
+
+        def deserialize_model_patcher(data: Any) -> Any:
+            """Deserialize ModelPatcher refs; pass through already-materialized objects."""
+            if isinstance(data, dict):
+                return ModelPatcherProxy(
+                    data["model_id"], registry=None, manage_lifecycle=False
+                )
+            return data
+
+        def deserialize_model_patcher_ref(data: Dict[str, Any]) -> Any:
+            """Context-aware ModelPatcherRef deserializer for both host and child."""
+            is_child = os.environ.get("PYISOLATE_CHILD") == "1"
+            if is_child:
+                return ModelPatcherProxy(
+                    data["model_id"], registry=None, manage_lifecycle=False
+                )
+            else:
+                return ModelPatcherRegistry()._get_instance(data["model_id"])
+
+        # Register ModelPatcher type for serialization
+        registry.register(
+            "ModelPatcher", serialize_model_patcher, deserialize_model_patcher
+        )
+        # Register ModelPatcherProxy type (already a proxy, just return ref)
+        registry.register(
+            "ModelPatcherProxy", serialize_model_patcher, deserialize_model_patcher
+        )
+        # Register ModelPatcherRef for deserialization (context-aware: host or child)
+        registry.register("ModelPatcherRef", None, deserialize_model_patcher_ref)
+
+        def serialize_clip(obj: Any) -> Dict[str, Any]:
+            if hasattr(obj, "_instance_id"):
+                return {"__type__": "CLIPRef", "clip_id": obj._instance_id}
+            clip_id = CLIPRegistry().register(obj)
+            return {"__type__": "CLIPRef", "clip_id": clip_id}
+
+        def deserialize_clip(data: Any) -> Any:
+            if isinstance(data, dict):
+                return CLIPProxy(data["clip_id"], registry=None, manage_lifecycle=False)
+            return data
+
+        def deserialize_clip_ref(data: Dict[str, Any]) -> Any:
+            """Context-aware CLIPRef deserializer for both host and child."""
+            is_child = os.environ.get("PYISOLATE_CHILD") == "1"
+            if is_child:
+                return CLIPProxy(data["clip_id"], registry=None, manage_lifecycle=False)
+            else:
+                return CLIPRegistry()._get_instance(data["clip_id"])
+
+        # Register CLIP type for serialization
+        registry.register("CLIP", serialize_clip, deserialize_clip)
+        # Register CLIPProxy type (already a proxy, just return ref)
+        registry.register("CLIPProxy", serialize_clip, deserialize_clip)
+        # Register CLIPRef for deserialization (context-aware: host or child)
+        registry.register("CLIPRef", None, deserialize_clip_ref)
+
+        def serialize_vae(obj: Any) -> Dict[str, Any]:
+            if hasattr(obj, "_instance_id"):
+                return {"__type__": "VAERef", "vae_id": obj._instance_id}
+            vae_id = VAERegistry().register(obj)
+            return {"__type__": "VAERef", "vae_id": vae_id}
+
+        def deserialize_vae(data: Any) -> Any:
+            if isinstance(data, dict):
+                return VAEProxy(data["vae_id"])
+            return data
+
+        def deserialize_vae_ref(data: Dict[str, Any]) -> Any:
+            """Context-aware VAERef deserializer for both host and child."""
+            is_child = os.environ.get("PYISOLATE_CHILD") == "1"
+            if is_child:
+                # Child: create a proxy
+                return VAEProxy(data["vae_id"])
+            else:
+                # Host: lookup real VAE from registry
+                return VAERegistry()._get_instance(data["vae_id"])
+
+        # Register VAE type for serialization
+        registry.register("VAE", serialize_vae, deserialize_vae)
+        # Register VAEProxy type (already a proxy, just return ref)
+        registry.register("VAEProxy", serialize_vae, deserialize_vae)
+        # Register VAERef for deserialization (context-aware: host or child)
+        registry.register("VAERef", None, deserialize_vae_ref)
+
+        # ModelSampling serialization - handles ModelSampling* types
+        # copyreg removed - no pickle fallback allowed
+
+        def serialize_model_sampling(obj: Any) -> Dict[str, Any]:
+            # Proxy with _instance_id — return ref (works from both host and child)
+            if hasattr(obj, "_instance_id"):
+                return {"__type__": "ModelSamplingRef", "ms_id": obj._instance_id}
+            # Child-side: object created locally in child (e.g. ModelSamplingAdvanced
+            # in nodes_z_image_turbo.py). Serialize as inline data so the host can
+            # reconstruct the real torch.nn.Module.
+            if os.environ.get("PYISOLATE_CHILD") == "1":
+                import base64
+                import io as _io
+
+                # Identify base classes from comfy.model_sampling
+                bases = []
+                for base in type(obj).__mro__:
+                    if base.__module__ == "comfy.model_sampling" and base.__name__ != "object":
+                        bases.append(base.__name__)
+                # Serialize state_dict as base64 safetensors-like
+                sd = obj.state_dict()
+                sd_serialized = {}
+                for k, v in sd.items():
+                    buf = _io.BytesIO()
+                    torch.save(v, buf)
+                    sd_serialized[k] = base64.b64encode(buf.getvalue()).decode("ascii")
+                # Capture plain attrs (shift, multiplier, sigma_data, etc.)
+                plain_attrs = {}
+                for k, v in obj.__dict__.items():
+                    if k.startswith("_"):
+                        continue
+                    if isinstance(v, (bool, int, float, str)):
+                        plain_attrs[k] = v
+                return {
+                    "__type__": "ModelSamplingInline",
+                    "bases": bases,
+                    "state_dict": sd_serialized,
+                    "attrs": plain_attrs,
+                }
+            # Host-side: register with ModelSamplingRegistry and return JSON-safe dict
+            ms_id = ModelSamplingRegistry().register(obj)
+            return {"__type__": "ModelSamplingRef", "ms_id": ms_id}
+
+        def deserialize_model_sampling(data: Any) -> Any:
+            """Deserialize ModelSampling refs or inline data."""
+            if isinstance(data, dict):
+                if data.get("__type__") == "ModelSamplingInline":
+                    return _reconstruct_model_sampling_inline(data)
+                return ModelSamplingProxy(data["ms_id"])
+            return data
+
+        def _reconstruct_model_sampling_inline(data: Dict[str, Any]) -> Any:
+            """Reconstruct a ModelSampling object on the host from inline child data."""
+            import comfy.model_sampling as _ms
+            import base64
+            import io as _io
+
+            # Resolve base classes
+            base_classes = []
+            for name in data["bases"]:
+                cls = getattr(_ms, name, None)
+                if cls is not None:
+                    base_classes.append(cls)
+            if not base_classes:
+                raise RuntimeError(
+                    f"Cannot reconstruct ModelSampling: no known bases in {data['bases']}"
+                )
+            # Create dynamic class matching the child's class hierarchy
+            ReconstructedSampling = type("ReconstructedSampling", tuple(base_classes), {})
+            obj = ReconstructedSampling.__new__(ReconstructedSampling)
+            torch.nn.Module.__init__(obj)
+            # Restore plain attributes first
+            for k, v in data.get("attrs", {}).items():
+                setattr(obj, k, v)
+            # Restore state_dict (buffers like sigmas)
+            for k, v_b64 in data.get("state_dict", {}).items():
+                buf = _io.BytesIO(base64.b64decode(v_b64))
+                tensor = torch.load(buf, weights_only=True)
+                # Register as buffer so it's part of state_dict
+                parts = k.split(".")
+                if len(parts) == 1:
+                    cast(Any, obj).register_buffer(parts[0], tensor)  # pylint: disable=no-member
+                else:
+                    setattr(obj, parts[0], tensor)
+            # Register on host so future references use proxy pattern.
+            # Skip in child process — register() is async RPC and cannot be
+            # called synchronously during deserialization.
+            if os.environ.get("PYISOLATE_CHILD") != "1":
+                ModelSamplingRegistry().register(obj)
+            return obj
+
+        def deserialize_model_sampling_ref(data: Dict[str, Any]) -> Any:
+            """Context-aware ModelSamplingRef deserializer for both host and child."""
+            is_child = os.environ.get("PYISOLATE_CHILD") == "1"
+            if is_child:
+                return ModelSamplingProxy(data["ms_id"])
+            else:
+                return ModelSamplingRegistry()._get_instance(data["ms_id"])
+
+        # Register all ModelSampling* and StableCascadeSampling classes dynamically
+        import comfy.model_sampling
+
+        for ms_cls in vars(comfy.model_sampling).values():
+            if not isinstance(ms_cls, type):
+                continue
+            if not issubclass(ms_cls, torch.nn.Module):
+                continue
+            if not (ms_cls.__name__.startswith("ModelSampling") or ms_cls.__name__ == "StableCascadeSampling"):
+                continue
+            registry.register(
+                ms_cls.__name__,
+                serialize_model_sampling,
+                deserialize_model_sampling,
+            )
+        registry.register(
+            "ModelSamplingProxy", serialize_model_sampling, deserialize_model_sampling
+        )
+        # Register ModelSamplingRef for deserialization (context-aware: host or child)
+        registry.register("ModelSamplingRef", None, deserialize_model_sampling_ref)
+        # Register ModelSamplingInline for deserialization (child→host inline transfer)
+        registry.register(
+            "ModelSamplingInline", None, lambda data: _reconstruct_model_sampling_inline(data)
+        )
+
+        def serialize_cond(obj: Any) -> Dict[str, Any]:
+            type_key = f"{type(obj).__module__}.{type(obj).__name__}"
+            return {
+                "__type__": type_key,
+                "cond": obj.cond,
+            }
+
+        def deserialize_cond(data: Dict[str, Any]) -> Any:
+            import importlib
+
+            type_key = data["__type__"]
+            module_name, class_name = type_key.rsplit(".", 1)
+            module = importlib.import_module(module_name)
+            cls = getattr(module, class_name)
+            return cls(data["cond"])
+
+        def _serialize_public_state(obj: Any) -> Dict[str, Any]:
+            state: Dict[str, Any] = {}
+            for key, value in obj.__dict__.items():
+                if key.startswith("_"):
+                    continue
+                if callable(value):
+                    continue
+                state[key] = value
+            return state
+
+        def serialize_latent_format(obj: Any) -> Dict[str, Any]:
+            type_key = f"{type(obj).__module__}.{type(obj).__name__}"
+            return {
+                "__type__": type_key,
+                "state": _serialize_public_state(obj),
+            }
+
+        def deserialize_latent_format(data: Dict[str, Any]) -> Any:
+            import importlib
+
+            type_key = data["__type__"]
+            module_name, class_name = type_key.rsplit(".", 1)
+            module = importlib.import_module(module_name)
+            cls = getattr(module, class_name)
+            obj = cls()
+            for key, value in data.get("state", {}).items():
+                prop = getattr(type(obj), key, None)
+                if isinstance(prop, property) and prop.fset is None:
+                    continue
+                setattr(obj, key, value)
+            return obj
+
+        import comfy.conds
+
+        for cond_cls in vars(comfy.conds).values():
+            if not isinstance(cond_cls, type):
+                continue
+            if not issubclass(cond_cls, comfy.conds.CONDRegular):
+                continue
+            type_key = f"{cond_cls.__module__}.{cond_cls.__name__}"
+            registry.register(type_key, serialize_cond, deserialize_cond)
+            registry.register(cond_cls.__name__, serialize_cond, deserialize_cond)
+
+        import comfy.latent_formats
+
+        for latent_cls in vars(comfy.latent_formats).values():
+            if not isinstance(latent_cls, type):
+                continue
+            if not issubclass(latent_cls, comfy.latent_formats.LatentFormat):
+                continue
+            type_key = f"{latent_cls.__module__}.{latent_cls.__name__}"
+            registry.register(
+                type_key, serialize_latent_format, deserialize_latent_format
+            )
+            registry.register(
+                latent_cls.__name__, serialize_latent_format, deserialize_latent_format
+            )
+
+        # V3 API: unwrap NodeOutput.args
+        def deserialize_node_output(data: Any) -> Any:
+            return getattr(data, "args", data)
+
+        registry.register("NodeOutput", None, deserialize_node_output)
+
+        # KSAMPLER serializer: stores sampler name instead of function object
+        # sampler_function is a callable which gets filtered out by JSONSocketTransport
+        def serialize_ksampler(obj: Any) -> Dict[str, Any]:
+            func_name = obj.sampler_function.__name__
+            # Map function name back to sampler name
+            if func_name == "sample_unipc":
+                sampler_name = "uni_pc"
+            elif func_name == "sample_unipc_bh2":
+                sampler_name = "uni_pc_bh2"
+            elif func_name == "dpm_fast_function":
+                sampler_name = "dpm_fast"
+            elif func_name == "dpm_adaptive_function":
+                sampler_name = "dpm_adaptive"
+            elif func_name.startswith("sample_"):
+                sampler_name = func_name[7:]  # Remove "sample_" prefix
+            else:
+                sampler_name = func_name
+            return {
+                "__type__": "KSAMPLER",
+                "sampler_name": sampler_name,
+                "extra_options": obj.extra_options,
+                "inpaint_options": obj.inpaint_options,
+            }
+
+        def deserialize_ksampler(data: Dict[str, Any]) -> Any:
+            import comfy.samplers
+
+            return comfy.samplers.ksampler(
+                data["sampler_name"],
+                data.get("extra_options", {}),
+                data.get("inpaint_options", {}),
+            )
+
+        registry.register("KSAMPLER", serialize_ksampler, deserialize_ksampler)
+
+        from comfy.isolation.model_patcher_proxy_utils import register_hooks_serializers
+
+        register_hooks_serializers(registry)
+
+        # -- File3D (comfy_api.latest._util.geometry_types) ---------------------
+        # Origin: comfy_api by ComfyOrg (Alexander Piskun), PR #12129
+
+        def serialize_file3d(obj: Any) -> Dict[str, Any]:
+            import base64
+            return {
+                "__type__": "File3D",
+                "format": obj.format,
+                "data": base64.b64encode(obj.get_bytes()).decode("ascii"),
+            }
+
+        def deserialize_file3d(data: Any) -> Any:
+            import base64
+            from io import BytesIO
+            from comfy_api.latest._util.geometry_types import File3D
+            return File3D(BytesIO(base64.b64decode(data["data"])), file_format=data["format"])
+
+        registry.register("File3D", serialize_file3d, deserialize_file3d, data_type=True)
+
+        # -- VIDEO (comfy_api.latest._input_impl.video_types) -------------------
+        # Origin: ComfyAPI Core v0.0.2 by ComfyOrg (guill), PR #8962
+
+        def serialize_video(obj: Any) -> Dict[str, Any]:
+            components = obj.get_components()
+            images = components.images.detach() if components.images.requires_grad else components.images
+            result: Dict[str, Any] = {
+                "__type__": "VIDEO",
+                "images": images,
+                "frame_rate_num": components.frame_rate.numerator,
+                "frame_rate_den": components.frame_rate.denominator,
+            }
+            if components.audio is not None:
+                waveform = components.audio["waveform"]
+                if waveform.requires_grad:
+                    waveform = waveform.detach()
+                result["audio_waveform"] = waveform
+                result["audio_sample_rate"] = components.audio["sample_rate"]
+            if components.metadata is not None:
+                result["metadata"] = components.metadata
+            return result
+
+        def deserialize_video(data: Any) -> Any:
+            from fractions import Fraction
+            from comfy_api.latest._input_impl.video_types import VideoFromComponents
+            from comfy_api.latest._util.video_types import VideoComponents
+            audio = None
+            if "audio_waveform" in data:
+                audio = {"waveform": data["audio_waveform"], "sample_rate": data["audio_sample_rate"]}
+            components = VideoComponents(
+                images=data["images"],
+                frame_rate=Fraction(data["frame_rate_num"], data["frame_rate_den"]),
+                audio=audio,
+                metadata=data.get("metadata"),
+            )
+            return VideoFromComponents(components)
+
+        registry.register("VIDEO", serialize_video, deserialize_video, data_type=True)
+        registry.register("VideoFromFile", serialize_video, deserialize_video, data_type=True)
+        registry.register("VideoFromComponents", serialize_video, deserialize_video, data_type=True)
+
+    def setup_web_directory(self, module: Any) -> None:
+        """Detect WEB_DIRECTORY on a module and populate/register it.
+
+        Called by the sealed worker after loading the node module.
+        Mirrors extension_wrapper.py:216-227 for host-coupled nodes.
+        Does NOT import extension_wrapper.py (it has `import torch` at module level).
+        """
+        import shutil
+
+        web_dir_attr = getattr(module, "WEB_DIRECTORY", None)
+        if web_dir_attr is None:
+            return
+
+        module_dir = os.path.dirname(os.path.abspath(module.__file__))
+        web_dir_path = os.path.abspath(os.path.join(module_dir, web_dir_attr))
+
+        # Read extension name from pyproject.toml
+        ext_name = os.path.basename(module_dir)
+        pyproject = os.path.join(module_dir, "pyproject.toml")
+        if os.path.exists(pyproject):
+            try:
+                import tomllib
+            except ImportError:
+                import tomli as tomllib  # type: ignore[no-redef]
+            try:
+                with open(pyproject, "rb") as f:
+                    data = tomllib.load(f)
+                name = data.get("project", {}).get("name")
+                if name:
+                    ext_name = name
+            except Exception:
+                pass
+
+        # Populate web dir if empty (mirrors _run_prestartup_web_copy)
+        if not (os.path.isdir(web_dir_path) and any(os.scandir(web_dir_path))):
+            os.makedirs(web_dir_path, exist_ok=True)
+
+            # Module-defined copy spec
+            copy_spec = getattr(module, "_PRESTARTUP_WEB_COPY", None)
+            if copy_spec is not None and callable(copy_spec):
+                try:
+                    copy_spec(web_dir_path)
+                except Exception as e:
+                    logger.warning("][ _PRESTARTUP_WEB_COPY failed: %s", e)
+
+            # Fallback: comfy_3d_viewers
+            try:
+                from comfy_3d_viewers import copy_viewer, VIEWER_FILES
+                for viewer in VIEWER_FILES:
+                    try:
+                        copy_viewer(viewer, web_dir_path)
+                    except Exception:
+                        pass
+            except ImportError:
+                pass
+
+            # Fallback: comfy_dynamic_widgets
+            try:
+                from comfy_dynamic_widgets import get_js_path
+                src = os.path.realpath(get_js_path())
+                if os.path.exists(src):
+                    dst_dir = os.path.join(web_dir_path, "js")
+                    os.makedirs(dst_dir, exist_ok=True)
+                    shutil.copy2(src, os.path.join(dst_dir, "dynamic_widgets.js"))
+            except ImportError:
+                pass
+
+        if os.path.isdir(web_dir_path) and any(os.scandir(web_dir_path)):
+            WebDirectoryProxy.register_web_dir(ext_name, web_dir_path)
+            logger.info(
+                "][ Adapter: registered web dir for %s (%d files)",
+                ext_name,
+                sum(1 for _ in Path(web_dir_path).rglob("*") if _.is_file()),
+            )
+
+    @staticmethod
+    def register_host_event_handlers(extension: Any) -> None:
+        """Register host-side event handlers for an isolated extension.
+
+        Wires ``"progress"`` events from the child to ``comfy.utils.PROGRESS_BAR_HOOK``
+        so the ComfyUI frontend receives progress bar updates.
+        """
+        register_event_handler = inspect.getattr_static(
+            extension, "register_event_handler", None
+        )
+        if not callable(register_event_handler):
+            return
+
+        def _host_progress_handler(payload: dict) -> None:
+            import comfy.utils
+
+            hook = comfy.utils.PROGRESS_BAR_HOOK
+            if hook is not None:
+                hook(
+                    payload.get("value", 0),
+                    payload.get("total", 0),
+                    payload.get("preview"),
+                    payload.get("node_id"),
+                )
+
+        extension.register_event_handler("progress", _host_progress_handler)
+
+    def setup_child_event_hooks(self, extension: Any) -> None:
+        """Wire PROGRESS_BAR_HOOK in the child to emit_event on the extension.
+
+        Host-coupled only — sealed workers do not have comfy.utils (torch).
+        """
+        is_child = os.environ.get("PYISOLATE_CHILD") == "1"
+        logger.info("][ ISO:setup_child_event_hooks called, PYISOLATE_CHILD=%s", is_child)
+        if not is_child:
+            return
+
+        if not _IMPORT_TORCH:
+            logger.info("][ ISO:setup_child_event_hooks skipped — sealed worker (no torch)")
+            return
+
+        import comfy.utils
+
+        def _event_progress_hook(value, total, preview=None, node_id=None):
+            logger.debug("][ ISO:event_progress value=%s/%s node_id=%s", value, total, node_id)
+            extension.emit_event("progress", {
+                "value": value,
+                "total": total,
+                "node_id": node_id,
+            })
+
+        comfy.utils.PROGRESS_BAR_HOOK = _event_progress_hook
+        logger.info("][ ISO:PROGRESS_BAR_HOOK wired to event channel")
+
+    def provide_rpc_services(self) -> List[type[ProxiedSingleton]]:
+        # Always available — no torch/PIL dependency
+        services: List[type[ProxiedSingleton]] = [
+            FolderPathsProxy,
+            HelperProxiesService,
+            WebDirectoryProxy,
+        ]
+        # Torch/PIL-dependent proxies
+        if _HAS_TORCH_PROXIES:
+            services.extend([
+                PromptServerService,
+                ModelManagementProxy,
+                UtilsProxy,
+                ProgressProxy,
+                VAERegistry,
+                CLIPRegistry,
+                ModelPatcherRegistry,
+                ModelSamplingRegistry,
+                FirstStageModelRegistry,
+            ])
+        return services
+
+    def handle_api_registration(self, api: ProxiedSingleton, rpc: AsyncRPC) -> None:
+        # Resolve the real name whether it's an instance or the Singleton class itself
+        api_name = api.__name__ if isinstance(api, type) else api.__class__.__name__
+
+        if api_name == "FolderPathsProxy":
+            import folder_paths
+
+            # Replace module-level functions with proxy methods
+            # This is aggressive but necessary for transparent proxying
+            # Handle both instance and class cases
+            instance = api() if isinstance(api, type) else api
+            for name in dir(instance):
+                if not name.startswith("_"):
+                    setattr(folder_paths, name, getattr(instance, name))
+
+            # Fence: isolated children get writable temp inside sandbox
+            if os.environ.get("PYISOLATE_CHILD") == "1":
+                import tempfile
+                _child_temp = os.path.join(tempfile.gettempdir(), "comfyui_temp")
+                os.makedirs(_child_temp, exist_ok=True)
+                folder_paths.temp_directory = _child_temp
+
+            return
+
+        if api_name == "ModelManagementProxy":
+            if _IMPORT_TORCH:
+                import comfy.model_management
+
+                instance = api() if isinstance(api, type) else api
+                # Replace module-level functions with proxy methods
+                for name in dir(instance):
+                    if not name.startswith("_"):
+                        setattr(comfy.model_management, name, getattr(instance, name))
+            return
+
+        if api_name == "UtilsProxy":
+            if not _IMPORT_TORCH:
+                logger.info("][ ISO:UtilsProxy handle_api_registration skipped — sealed worker (no torch)")
+                return
+
+            import comfy.utils
+
+            # Static Injection of RPC mechanism to ensure Child can access it
+            # independent of instance lifecycle.
+            api.set_rpc(rpc)
+
+            is_child = os.environ.get("PYISOLATE_CHILD") == "1"
+            logger.info("][ ISO:UtilsProxy handle_api_registration PYISOLATE_CHILD=%s", is_child)
+
+            # Progress hook wiring moved to setup_child_event_hooks via event channel
+
+            return
+
+        if api_name == "PromptServerService":
+            if not _IMPORT_TORCH:
+                return
+            import server
+            from comfy.isolation.proxies.prompt_server_impl import PromptServerStub
+
+            stub = PromptServerStub()
+            if (
+                hasattr(server, "PromptServer")
+                and getattr(server.PromptServer, "instance", None) is not stub
+            ):
+                server.PromptServer.instance = stub

comfy/isolation/child_hooks.py

@@ -0,0 +1,122 @@
+# pylint: disable=import-outside-toplevel,logging-fstring-interpolation
+# Child process initialization for PyIsolate
+import logging
+import os
+
+logger = logging.getLogger(__name__)
+
+
+def is_child_process() -> bool:
+    return os.environ.get("PYISOLATE_CHILD") == "1"
+
+
+def _load_extra_model_paths() -> None:
+    """Load extra_model_paths.yaml so the child's folder_paths has the same search paths as the host.
+
+    The host loads this in main.py:143-145. The child is spawned by
+    pyisolate's uds_client.py and never runs main.py, so folder_paths
+    only has the base model directories. Any isolated node calling
+    folder_paths.get_filename_list() in define_schema() would get empty
+    results for folders whose files live in extra_model_paths locations.
+    """
+    import folder_paths  # noqa: F401 — side-effect import; load_extra_path_config writes to folder_paths internals
+    from utils.extra_config import load_extra_path_config
+
+    extra_config_path = os.path.join(
+        os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))),
+        "extra_model_paths.yaml",
+    )
+    if os.path.isfile(extra_config_path):
+        load_extra_path_config(extra_config_path)
+
+
+def initialize_child_process() -> None:
+    if os.environ.get("PYISOLATE_IMPORT_TORCH", "1") != "0":
+        _load_extra_model_paths()
+    _setup_child_loop_bridge()
+
+    # Manual RPC injection
+    try:
+        from pyisolate._internal.rpc_protocol import get_child_rpc_instance
+
+        rpc = get_child_rpc_instance()
+        if rpc:
+            _setup_proxy_callers(rpc)
+        else:
+            _setup_proxy_callers()
+    except Exception as e:
+        logger.error(f"][ child_hooks Manual RPC Injection failed: {e}")
+        _setup_proxy_callers()
+
+    _setup_logging()
+
+
+def _setup_child_loop_bridge() -> None:
+    import asyncio
+
+    main_loop = None
+    try:
+        main_loop = asyncio.get_running_loop()
+    except RuntimeError:
+        try:
+            main_loop = asyncio.get_event_loop()
+        except RuntimeError:
+            pass
+
+    if main_loop is None:
+        return
+
+    try:
+        from .proxies.base import set_global_loop
+
+        set_global_loop(main_loop)
+    except ImportError:
+        pass
+
+
+def _setup_prompt_server_stub(rpc=None) -> None:
+    try:
+        from .proxies.prompt_server_impl import PromptServerStub
+
+        if rpc:
+            PromptServerStub.set_rpc(rpc)
+        elif hasattr(PromptServerStub, "clear_rpc"):
+            PromptServerStub.clear_rpc()
+        else:
+            PromptServerStub._rpc = None  # type: ignore[attr-defined]
+
+    except Exception as e:
+        logger.error(f"Failed to setup PromptServerStub: {e}")
+
+
+def _setup_proxy_callers(rpc=None) -> None:
+    try:
+        from .proxies.folder_paths_proxy import FolderPathsProxy
+        from .proxies.helper_proxies import HelperProxiesService
+        from .proxies.model_management_proxy import ModelManagementProxy
+        from .proxies.progress_proxy import ProgressProxy
+        from .proxies.prompt_server_impl import PromptServerStub
+        from .proxies.utils_proxy import UtilsProxy
+
+        if rpc is None:
+            FolderPathsProxy.clear_rpc()
+            HelperProxiesService.clear_rpc()
+            ModelManagementProxy.clear_rpc()
+            ProgressProxy.clear_rpc()
+            PromptServerStub.clear_rpc()
+            UtilsProxy.clear_rpc()
+            return
+
+        FolderPathsProxy.set_rpc(rpc)
+        HelperProxiesService.set_rpc(rpc)
+        ModelManagementProxy.set_rpc(rpc)
+        ProgressProxy.set_rpc(rpc)
+        PromptServerStub.set_rpc(rpc)
+        UtilsProxy.set_rpc(rpc)
+
+    except Exception as e:
+        logger.error(f"Failed to setup child singleton proxy callers: {e}")
+
+
+def _setup_logging() -> None:
+    logging.getLogger().setLevel(logging.INFO)

comfy/isolation/clip_proxy.py

@@ -0,0 +1,327 @@
+# pylint: disable=attribute-defined-outside-init,import-outside-toplevel,logging-fstring-interpolation
+# CLIP Proxy implementation
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING, Any, Optional
+
+from comfy.isolation.proxies.base import (
+    IS_CHILD_PROCESS,
+    BaseProxy,
+    BaseRegistry,
+    detach_if_grad,
+)
+
+if TYPE_CHECKING:
+    from comfy.isolation.model_patcher_proxy import ModelPatcherProxy
+
+
+class CondStageModelRegistry(BaseRegistry[Any]):
+    _type_prefix = "cond_stage_model"
+
+    async def get_property(self, instance_id: str, name: str) -> Any:
+        obj = self._get_instance(instance_id)
+        return getattr(obj, name)
+
+
+class CondStageModelProxy(BaseProxy[CondStageModelRegistry]):
+    _registry_class = CondStageModelRegistry
+    __module__ = "comfy.sd"
+
+    def __getattr__(self, name: str) -> Any:
+        try:
+            return self._call_rpc("get_property", name)
+        except Exception as e:
+            raise AttributeError(
+                f"'{self.__class__.__name__}' object has no attribute '{name}'"
+            ) from e
+
+    def __repr__(self) -> str:
+        return f"<CondStageModelProxy {self._instance_id}>"
+
+
+class TokenizerRegistry(BaseRegistry[Any]):
+    _type_prefix = "tokenizer"
+
+    async def get_property(self, instance_id: str, name: str) -> Any:
+        obj = self._get_instance(instance_id)
+        return getattr(obj, name)
+
+
+class TokenizerProxy(BaseProxy[TokenizerRegistry]):
+    _registry_class = TokenizerRegistry
+    __module__ = "comfy.sd"
+
+    def __getattr__(self, name: str) -> Any:
+        try:
+            return self._call_rpc("get_property", name)
+        except Exception as e:
+            raise AttributeError(
+                f"'{self.__class__.__name__}' object has no attribute '{name}'"
+            ) from e
+
+    def __repr__(self) -> str:
+        return f"<TokenizerProxy {self._instance_id}>"
+
+
+logger = logging.getLogger(__name__)
+
+
+class CLIPRegistry(BaseRegistry[Any]):
+    _type_prefix = "clip"
+    _allowed_setters = {
+        "layer_idx",
+        "tokenizer_options",
+        "use_clip_schedule",
+        "apply_hooks_to_conds",
+    }
+
+    async def get_ram_usage(self, instance_id: str) -> int:
+        return self._get_instance(instance_id).get_ram_usage()
+
+    async def get_patcher_id(self, instance_id: str) -> str:
+        from comfy.isolation.model_patcher_proxy import ModelPatcherRegistry
+
+        return ModelPatcherRegistry().register(self._get_instance(instance_id).patcher)
+
+    async def get_cond_stage_model_id(self, instance_id: str) -> str:
+        return CondStageModelRegistry().register(
+            self._get_instance(instance_id).cond_stage_model
+        )
+
+    async def get_tokenizer_id(self, instance_id: str) -> str:
+        return TokenizerRegistry().register(self._get_instance(instance_id).tokenizer)
+
+    async def load_model(self, instance_id: str) -> None:
+        self._get_instance(instance_id).load_model()
+
+    async def clip_layer(self, instance_id: str, layer_idx: int) -> None:
+        self._get_instance(instance_id).clip_layer(layer_idx)
+
+    async def set_tokenizer_option(
+        self, instance_id: str, option_name: str, value: Any
+    ) -> None:
+        self._get_instance(instance_id).set_tokenizer_option(option_name, value)
+
+    async def get_property(self, instance_id: str, name: str) -> Any:
+        return getattr(self._get_instance(instance_id), name)
+
+    async def set_property(self, instance_id: str, name: str, value: Any) -> None:
+        if name not in self._allowed_setters:
+            raise PermissionError(f"Setting '{name}' is not allowed via RPC")
+        setattr(self._get_instance(instance_id), name, value)
+
+    async def tokenize(
+        self, instance_id: str, text: str, return_word_ids: bool = False, **kwargs: Any
+    ) -> Any:
+        return self._get_instance(instance_id).tokenize(
+            text, return_word_ids=return_word_ids, **kwargs
+        )
+
+    async def encode(self, instance_id: str, text: str) -> Any:
+        return detach_if_grad(self._get_instance(instance_id).encode(text))
+
+    async def encode_from_tokens(
+        self,
+        instance_id: str,
+        tokens: Any,
+        return_pooled: bool = False,
+        return_dict: bool = False,
+    ) -> Any:
+        return detach_if_grad(
+            self._get_instance(instance_id).encode_from_tokens(
+                tokens, return_pooled=return_pooled, return_dict=return_dict
+            )
+        )
+
+    async def encode_from_tokens_scheduled(
+        self,
+        instance_id: str,
+        tokens: Any,
+        unprojected: bool = False,
+        add_dict: Optional[dict] = None,
+        show_pbar: bool = True,
+    ) -> Any:
+        add_dict = add_dict or {}
+        return detach_if_grad(
+            self._get_instance(instance_id).encode_from_tokens_scheduled(
+                tokens, unprojected=unprojected, add_dict=add_dict, show_pbar=show_pbar
+            )
+        )
+
+    async def add_patches(
+        self,
+        instance_id: str,
+        patches: Any,
+        strength_patch: float = 1.0,
+        strength_model: float = 1.0,
+    ) -> Any:
+        return self._get_instance(instance_id).add_patches(
+            patches, strength_patch=strength_patch, strength_model=strength_model
+        )
+
+    async def get_key_patches(self, instance_id: str) -> Any:
+        return self._get_instance(instance_id).get_key_patches()
+
+    async def load_sd(
+        self, instance_id: str, sd: dict, full_model: bool = False
+    ) -> Any:
+        return self._get_instance(instance_id).load_sd(sd, full_model=full_model)
+
+    async def get_sd(self, instance_id: str) -> Any:
+        return self._get_instance(instance_id).get_sd()
+
+    async def clone(self, instance_id: str) -> str:
+        return self.register(self._get_instance(instance_id).clone())
+
+
+class CLIPProxy(BaseProxy[CLIPRegistry]):
+    _registry_class = CLIPRegistry
+    __module__ = "comfy.sd"
+
+    def get_ram_usage(self) -> int:
+        return self._call_rpc("get_ram_usage")
+
+    @property
+    def patcher(self) -> "ModelPatcherProxy":
+        from comfy.isolation.model_patcher_proxy import ModelPatcherProxy
+
+        if not hasattr(self, "_patcher_proxy"):
+            patcher_id = self._call_rpc("get_patcher_id")
+            self._patcher_proxy = ModelPatcherProxy(patcher_id, manage_lifecycle=False)
+        return self._patcher_proxy
+
+    @patcher.setter
+    def patcher(self, value: Any) -> None:
+        from comfy.isolation.model_patcher_proxy import ModelPatcherProxy
+
+        if isinstance(value, ModelPatcherProxy):
+            self._patcher_proxy = value
+        else:
+            logger.warning(
+                f"Attempted to set CLIPProxy.patcher to non-proxy object: {value}"
+            )
+
+    @property
+    def cond_stage_model(self) -> CondStageModelProxy:
+        if not hasattr(self, "_cond_stage_model_proxy"):
+            csm_id = self._call_rpc("get_cond_stage_model_id")
+            self._cond_stage_model_proxy = CondStageModelProxy(
+                csm_id, manage_lifecycle=False
+            )
+        return self._cond_stage_model_proxy
+
+    @property
+    def tokenizer(self) -> TokenizerProxy:
+        if not hasattr(self, "_tokenizer_proxy"):
+            tok_id = self._call_rpc("get_tokenizer_id")
+            self._tokenizer_proxy = TokenizerProxy(tok_id, manage_lifecycle=False)
+        return self._tokenizer_proxy
+
+    def load_model(self) -> ModelPatcherProxy:
+        self._call_rpc("load_model")
+        return self.patcher
+
+    @property
+    def layer_idx(self) -> Optional[int]:
+        return self._call_rpc("get_property", "layer_idx")
+
+    @layer_idx.setter
+    def layer_idx(self, value: Optional[int]) -> None:
+        self._call_rpc("set_property", "layer_idx", value)
+
+    @property
+    def tokenizer_options(self) -> dict:
+        return self._call_rpc("get_property", "tokenizer_options")
+
+    @tokenizer_options.setter
+    def tokenizer_options(self, value: dict) -> None:
+        self._call_rpc("set_property", "tokenizer_options", value)
+
+    @property
+    def use_clip_schedule(self) -> bool:
+        return self._call_rpc("get_property", "use_clip_schedule")
+
+    @use_clip_schedule.setter
+    def use_clip_schedule(self, value: bool) -> None:
+        self._call_rpc("set_property", "use_clip_schedule", value)
+
+    @property
+    def apply_hooks_to_conds(self) -> Any:
+        return self._call_rpc("get_property", "apply_hooks_to_conds")
+
+    @apply_hooks_to_conds.setter
+    def apply_hooks_to_conds(self, value: Any) -> None:
+        self._call_rpc("set_property", "apply_hooks_to_conds", value)
+
+    def clip_layer(self, layer_idx: int) -> None:
+        return self._call_rpc("clip_layer", layer_idx)
+
+    def set_tokenizer_option(self, option_name: str, value: Any) -> None:
+        return self._call_rpc("set_tokenizer_option", option_name, value)
+
+    def tokenize(self, text: str, return_word_ids: bool = False, **kwargs: Any) -> Any:
+        return self._call_rpc(
+            "tokenize", text, return_word_ids=return_word_ids, **kwargs
+        )
+
+    def encode(self, text: str) -> Any:
+        return self._call_rpc("encode", text)
+
+    def encode_from_tokens(
+        self, tokens: Any, return_pooled: bool = False, return_dict: bool = False
+    ) -> Any:
+        res = self._call_rpc(
+            "encode_from_tokens",
+            tokens,
+            return_pooled=return_pooled,
+            return_dict=return_dict,
+        )
+        if return_pooled and isinstance(res, list) and not return_dict:
+            return tuple(res)
+        return res
+
+    def encode_from_tokens_scheduled(
+        self,
+        tokens: Any,
+        unprojected: bool = False,
+        add_dict: Optional[dict] = None,
+        show_pbar: bool = True,
+    ) -> Any:
+        add_dict = add_dict or {}
+        return self._call_rpc(
+            "encode_from_tokens_scheduled",
+            tokens,
+            unprojected=unprojected,
+            add_dict=add_dict,
+            show_pbar=show_pbar,
+        )
+
+    def add_patches(
+        self, patches: Any, strength_patch: float = 1.0, strength_model: float = 1.0
+    ) -> Any:
+        return self._call_rpc(
+            "add_patches",
+            patches,
+            strength_patch=strength_patch,
+            strength_model=strength_model,
+        )
+
+    def get_key_patches(self) -> Any:
+        return self._call_rpc("get_key_patches")
+
+    def load_sd(self, sd: dict, full_model: bool = False) -> Any:
+        return self._call_rpc("load_sd", sd, full_model=full_model)
+
+    def get_sd(self) -> Any:
+        return self._call_rpc("get_sd")
+
+    def clone(self) -> CLIPProxy:
+        new_id = self._call_rpc("clone")
+        return CLIPProxy(new_id, self._registry, manage_lifecycle=not IS_CHILD_PROCESS)
+
+
+if not IS_CHILD_PROCESS:
+    _CLIP_REGISTRY_SINGLETON = CLIPRegistry()
+    _COND_STAGE_MODEL_REGISTRY_SINGLETON = CondStageModelRegistry()
+    _TOKENIZER_REGISTRY_SINGLETON = TokenizerRegistry()

comfy/isolation/custom_node_serializers.py

@@ -0,0 +1,16 @@
+"""Compatibility shim for the indexed serializer path."""
+
+from __future__ import annotations
+
+from typing import Any
+
+
+def register_custom_node_serializers(_registry: Any) -> None:
+    """Legacy no-op shim.
+
+    Serializer registration now lives directly in the active isolation adapter.
+    This module remains importable because the isolation index still references it.
+    """
+    return None
+
+__all__ = ["register_custom_node_serializers"]

comfy/isolation/extension_loader.py

@@ -0,0 +1,540 @@
+# pylint: disable=cyclic-import,import-outside-toplevel,redefined-outer-name
+from __future__ import annotations
+
+import logging
+import os
+import inspect
+import sys
+import types
+import platform
+from pathlib import Path
+from typing import Any, Callable, Dict, List, Tuple
+
+import pyisolate
+from pyisolate import ExtensionManager, ExtensionManagerConfig
+from packaging.requirements import InvalidRequirement, Requirement
+from packaging.utils import canonicalize_name
+
+from .manifest_loader import is_cache_valid, load_from_cache, save_to_cache
+from .host_policy import load_host_policy
+
+try:
+    import tomllib
+except ImportError:
+    import tomli as tomllib  # type: ignore[no-redef]
+
+logger = logging.getLogger(__name__)
+
+
+def _register_web_directory(extension_name: str, node_dir: Path) -> None:
+    """Register an isolated extension's web directory on the host side."""
+    import nodes
+
+    # Method 1: pyproject.toml [tool.comfy] web field
+    pyproject = node_dir / "pyproject.toml"
+    if pyproject.exists():
+        try:
+            with pyproject.open("rb") as f:
+                data = tomllib.load(f)
+            web_dir_name = data.get("tool", {}).get("comfy", {}).get("web")
+            if web_dir_name:
+                web_dir_path = str(node_dir / web_dir_name)
+                if os.path.isdir(web_dir_path):
+                    nodes.EXTENSION_WEB_DIRS[extension_name] = web_dir_path
+                    logger.debug(
+                        "][ Registered web dir for isolated %s: %s",
+                        extension_name,
+                        web_dir_path,
+                    )
+                    return
+        except Exception:
+            pass
+
+    # Method 2: __init__.py WEB_DIRECTORY constant (parse without importing)
+    init_file = node_dir / "__init__.py"
+    if init_file.exists():
+        try:
+            source = init_file.read_text()
+            for line in source.splitlines():
+                stripped = line.strip()
+                if stripped.startswith("WEB_DIRECTORY"):
+                    # Parse: WEB_DIRECTORY = "./web" or WEB_DIRECTORY = "web"
+                    _, _, value = stripped.partition("=")
+                    value = value.strip().strip("\"'")
+                    if value:
+                        web_dir_path = str((node_dir / value).resolve())
+                        if os.path.isdir(web_dir_path):
+                            nodes.EXTENSION_WEB_DIRS[extension_name] = web_dir_path
+                            logger.debug(
+                                "][ Registered web dir for isolated %s: %s",
+                                extension_name,
+                                web_dir_path,
+                            )
+                            return
+        except Exception:
+            pass
+
+
+def _get_extension_type(execution_model: str) -> type[Any]:
+    if execution_model == "sealed_worker":
+        return pyisolate.SealedNodeExtension
+
+    from .extension_wrapper import ComfyNodeExtension
+
+    return ComfyNodeExtension
+
+
+async def _stop_extension_safe(extension: Any, extension_name: str) -> None:
+    try:
+        stop_result = extension.stop()
+        if inspect.isawaitable(stop_result):
+            await stop_result
+    except Exception:
+        logger.debug("][ %s stop failed", extension_name, exc_info=True)
+
+
+def _normalize_dependency_spec(dep: str, base_paths: list[Path]) -> str:
+    req, sep, marker = dep.partition(";")
+    req = req.strip()
+    marker_suffix = f";{marker}" if sep else ""
+
+    def _resolve_local_path(local_path: str) -> Path | None:
+        for base in base_paths:
+            candidate = (base / local_path).resolve()
+            if candidate.exists():
+                return candidate
+        return None
+
+    if req.startswith("./") or req.startswith("../"):
+        resolved = _resolve_local_path(req)
+        if resolved is not None:
+            return f"{resolved}{marker_suffix}"
+
+    if req.startswith("file://"):
+        raw = req[len("file://") :]
+        if raw.startswith("./") or raw.startswith("../"):
+            resolved = _resolve_local_path(raw)
+            if resolved is not None:
+                return f"file://{resolved}{marker_suffix}"
+
+    return dep
+
+
+def _dependency_name_from_spec(dep: str) -> str | None:
+    stripped = dep.strip()
+    if not stripped or stripped == "-e" or stripped.startswith("-e "):
+        return None
+    if stripped.startswith(("/", "./", "../", "file://")):
+        return None
+
+    try:
+        return canonicalize_name(Requirement(stripped).name)
+    except InvalidRequirement:
+        return None
+
+
+def _parse_cuda_wheels_config(
+    tool_config: dict[str, object], dependencies: list[str]
+) -> dict[str, object] | None:
+    raw_config = tool_config.get("cuda_wheels")
+    if raw_config is None:
+        return None
+    if not isinstance(raw_config, dict):
+        raise ExtensionLoadError("[tool.comfy.isolation.cuda_wheels] must be a table")
+
+    index_url = raw_config.get("index_url")
+    index_urls = raw_config.get("index_urls")
+    if index_urls is not None:
+        if not isinstance(index_urls, list) or not all(
+            isinstance(u, str) and u.strip() for u in index_urls
+        ):
+            raise ExtensionLoadError(
+                "[tool.comfy.isolation.cuda_wheels.index_urls] must be a list of non-empty strings"
+            )
+    elif not isinstance(index_url, str) or not index_url.strip():
+        raise ExtensionLoadError(
+            "[tool.comfy.isolation.cuda_wheels.index_url] must be a non-empty string"
+        )
+
+    packages = raw_config.get("packages")
+    if not isinstance(packages, list) or not all(
+        isinstance(package_name, str) and package_name.strip()
+        for package_name in packages
+    ):
+        raise ExtensionLoadError(
+            "[tool.comfy.isolation.cuda_wheels.packages] must be a list of non-empty strings"
+        )
+
+    declared_dependencies = {
+        dependency_name
+        for dep in dependencies
+        if (dependency_name := _dependency_name_from_spec(dep)) is not None
+    }
+    normalized_packages = [canonicalize_name(package_name) for package_name in packages]
+    missing = [
+        package_name
+        for package_name in normalized_packages
+        if package_name not in declared_dependencies
+    ]
+    if missing:
+        missing_joined = ", ".join(sorted(missing))
+        raise ExtensionLoadError(
+            "[tool.comfy.isolation.cuda_wheels.packages] references undeclared dependencies: "
+            f"{missing_joined}"
+        )
+
+    package_map = raw_config.get("package_map", {})
+    if not isinstance(package_map, dict):
+        raise ExtensionLoadError(
+            "[tool.comfy.isolation.cuda_wheels.package_map] must be a table"
+        )
+
+    normalized_package_map: dict[str, str] = {}
+    for dependency_name, index_package_name in package_map.items():
+        if not isinstance(dependency_name, str) or not dependency_name.strip():
+            raise ExtensionLoadError(
+                "[tool.comfy.isolation.cuda_wheels.package_map] keys must be non-empty strings"
+            )
+        if not isinstance(index_package_name, str) or not index_package_name.strip():
+            raise ExtensionLoadError(
+                "[tool.comfy.isolation.cuda_wheels.package_map] values must be non-empty strings"
+            )
+        canonical_dependency_name = canonicalize_name(dependency_name)
+        if canonical_dependency_name not in normalized_packages:
+            raise ExtensionLoadError(
+                "[tool.comfy.isolation.cuda_wheels.package_map] can only override packages listed in "
+                "[tool.comfy.isolation.cuda_wheels.packages]"
+            )
+        normalized_package_map[canonical_dependency_name] = index_package_name.strip()
+
+    result: dict = {
+        "packages": normalized_packages,
+        "package_map": normalized_package_map,
+    }
+    if index_urls is not None:
+        result["index_urls"] = [u.rstrip("/") + "/" for u in index_urls]
+    else:
+        result["index_url"] = index_url.rstrip("/") + "/"
+    return result
+
+
+def get_enforcement_policy() -> Dict[str, bool]:
+    return {
+        "force_isolated": os.environ.get("PYISOLATE_ENFORCE_ISOLATED") == "1",
+        "force_sandbox": os.environ.get("PYISOLATE_ENFORCE_SANDBOX") == "1",
+    }
+
+
+class ExtensionLoadError(RuntimeError):
+    pass
+
+
+def register_dummy_module(extension_name: str, node_dir: Path) -> None:
+    normalized_name = extension_name.replace("-", "_").replace(".", "_")
+    if normalized_name not in sys.modules:
+        dummy_module = types.ModuleType(normalized_name)
+        dummy_module.__file__ = str(node_dir / "__init__.py")
+        dummy_module.__path__ = [str(node_dir)]
+        dummy_module.__package__ = normalized_name
+        sys.modules[normalized_name] = dummy_module
+
+
+def _is_stale_node_cache(cached_data: Dict[str, Dict]) -> bool:
+    for details in cached_data.values():
+        if not isinstance(details, dict):
+            return True
+        if details.get("is_v3") and "schema_v1" not in details:
+            return True
+    return False
+
+
+async def load_isolated_node(
+    node_dir: Path,
+    manifest_path: Path,
+    logger: logging.Logger,
+    build_stub_class: Callable[[str, Dict[str, object], Any], type],
+    venv_root: Path,
+    extension_managers: List[ExtensionManager],
+) -> List[Tuple[str, str, type]]:
+    try:
+        with manifest_path.open("rb") as handle:
+            manifest_data = tomllib.load(handle)
+    except Exception as e:
+        logger.warning(f"][ Failed to parse {manifest_path}: {e}")
+        return []
+
+    # Parse [tool.comfy.isolation]
+    tool_config = manifest_data.get("tool", {}).get("comfy", {}).get("isolation", {})
+    can_isolate = tool_config.get("can_isolate", False)
+    share_torch = tool_config.get("share_torch", False)
+    package_manager = tool_config.get("package_manager", "uv")
+    is_conda = package_manager == "conda"
+    execution_model = tool_config.get("execution_model")
+    if execution_model is None:
+        execution_model = "sealed_worker" if is_conda else "host-coupled"
+
+    if "sealed_host_ro_paths" in tool_config:
+        raise ValueError(
+            "Manifest field 'sealed_host_ro_paths' is not allowed. "
+            "Configure [tool.comfy.host].sealed_worker_ro_import_paths in host policy."
+        )
+
+    # Conda-specific manifest fields
+    conda_channels: list[str] = (
+        tool_config.get("conda_channels", []) if is_conda else []
+    )
+    conda_dependencies: list[str] = (
+        tool_config.get("conda_dependencies", []) if is_conda else []
+    )
+    conda_platforms: list[str] = (
+        tool_config.get("conda_platforms", []) if is_conda else []
+    )
+    conda_python: str = (
+        tool_config.get("conda_python", "*") if is_conda else "*"
+    )
+
+    # Parse [project] dependencies
+    project_config = manifest_data.get("project", {})
+    dependencies = project_config.get("dependencies", [])
+    if not isinstance(dependencies, list):
+        dependencies = []
+
+    # Get extension name (default to folder name if not in project.name)
+    extension_name = project_config.get("name", node_dir.name)
+
+    # LOGIC: Isolation Decision
+    policy = get_enforcement_policy()
+    isolated = can_isolate or policy["force_isolated"]
+
+    if not isolated:
+        return []
+
+    import folder_paths
+
+    base_paths = [Path(folder_paths.base_path), node_dir]
+    dependencies = [
+        _normalize_dependency_spec(dep, base_paths) if isinstance(dep, str) else dep
+        for dep in dependencies
+    ]
+    cuda_wheels = _parse_cuda_wheels_config(tool_config, dependencies)
+
+    manager_config = ExtensionManagerConfig(venv_root_path=str(venv_root))
+    extension_type = _get_extension_type(execution_model)
+    manager: ExtensionManager = pyisolate.ExtensionManager(
+        extension_type, manager_config
+    )
+    extension_managers.append(manager)
+
+    host_policy = load_host_policy(Path(folder_paths.base_path))
+
+    sandbox_config = {}
+    is_linux = platform.system() == "Linux"
+
+    if is_conda:
+        share_torch = False
+        share_cuda_ipc = False
+    else:
+        share_cuda_ipc = share_torch and is_linux
+
+    if is_linux and isolated:
+        sandbox_config = {
+            "network": host_policy["allow_network"],
+            "writable_paths": host_policy["writable_paths"],
+            "readonly_paths": host_policy["readonly_paths"],
+        }
+
+    extension_config: dict = {
+        "name": extension_name,
+        "module_path": str(node_dir),
+        "isolated": True,
+        "dependencies": dependencies,
+        "share_torch": share_torch,
+        "share_cuda_ipc": share_cuda_ipc,
+        "sandbox_mode": host_policy["sandbox_mode"],
+        "sandbox": sandbox_config,
+    }
+
+    share_torch_no_deps = tool_config.get("share_torch_no_deps", [])
+    if share_torch_no_deps:
+        if not isinstance(share_torch_no_deps, list) or not all(
+            isinstance(dep, str) and dep.strip() for dep in share_torch_no_deps
+        ):
+            raise ExtensionLoadError(
+                "[tool.comfy.isolation.share_torch_no_deps] must be a list of non-empty strings"
+            )
+        extension_config["share_torch_no_deps"] = share_torch_no_deps
+
+    _is_sealed = execution_model == "sealed_worker"
+    _is_sandboxed = host_policy["sandbox_mode"] != "disabled" and is_linux
+    logger.info(
+        "][ Loading isolated node: %s (torch_share [%s], sealed [%s], sandboxed [%s])",
+        extension_name,
+        "x" if share_torch else " ",
+        "x" if _is_sealed else " ",
+        "x" if _is_sandboxed else " ",
+    )
+
+    if cuda_wheels is not None:
+        extension_config["cuda_wheels"] = cuda_wheels
+
+    extra_index_urls = tool_config.get("extra_index_urls", [])
+    if extra_index_urls:
+        if not isinstance(extra_index_urls, list) or not all(
+            isinstance(u, str) and u.strip() for u in extra_index_urls
+        ):
+            raise ExtensionLoadError(
+                "[tool.comfy.isolation.extra_index_urls] must be a list of non-empty strings"
+            )
+        extension_config["extra_index_urls"] = extra_index_urls
+
+    # Conda-specific keys
+    if is_conda:
+        extension_config["package_manager"] = "conda"
+        extension_config["conda_channels"] = conda_channels
+        extension_config["conda_dependencies"] = conda_dependencies
+        extension_config["conda_python"] = conda_python
+        find_links = tool_config.get("find_links", [])
+        if find_links:
+            extension_config["find_links"] = find_links
+        if conda_platforms:
+            extension_config["conda_platforms"] = conda_platforms
+
+    if execution_model != "host-coupled":
+        extension_config["execution_model"] = execution_model
+    if execution_model == "sealed_worker":
+        policy_ro_paths = host_policy.get("sealed_worker_ro_import_paths", [])
+        if isinstance(policy_ro_paths, list) and policy_ro_paths:
+            extension_config["sealed_host_ro_paths"] = list(policy_ro_paths)
+        # Sealed workers keep the host RPC service inventory even when the
+        # child resolves no API classes locally.
+
+    extension = manager.load_extension(extension_config)
+    register_dummy_module(extension_name, node_dir)
+
+    # Register host-side event handlers via adapter
+    from .adapter import ComfyUIAdapter
+    ComfyUIAdapter.register_host_event_handlers(extension)
+
+    # Register web directory on the host — only when sandbox is disabled.
+    # In sandbox mode, serving untrusted JS to the browser is not safe.
+    if host_policy["sandbox_mode"] == "disabled":
+        _register_web_directory(extension_name, node_dir)
+
+    # Register for proxied web serving — the child's web dir may have
+    # content that doesn't exist on the host (e.g., pip-installed viewer
+    # bundles). The WebDirectoryCache will lazily fetch via RPC.
+    from .proxies.web_directory_proxy import WebDirectoryProxy, get_web_directory_cache
+
+    class ChildWebDirectoryProxy:
+        def __init__(self, host_extension):
+            self._host_extension = host_extension
+            self._caller = None
+
+        def _get_caller(self):
+            self._host_extension.proxy
+            rpc = self._host_extension._extension.rpc
+            caller = rpc.create_caller(WebDirectoryProxy, WebDirectoryProxy.get_remote_id())
+            if self._caller is not caller:
+                self._caller = caller
+            return self._caller
+
+        def list_web_files(self, extension_name: str):
+            from .proxies.base import run_sync_rpc_coro
+            return run_sync_rpc_coro(self._get_caller().list_web_files(extension_name))
+
+        def get_web_file(self, extension_name: str, relative_path: str):
+            from .proxies.base import run_sync_rpc_coro
+            return run_sync_rpc_coro(
+                self._get_caller().get_web_file(extension_name, relative_path)
+            )
+
+    cache = get_web_directory_cache()
+    cache.register_proxy(extension_name, ChildWebDirectoryProxy(extension))
+
+    # Try cache first (lazy spawn)
+    if is_cache_valid(node_dir, manifest_path, venv_root):
+        cached_data = load_from_cache(node_dir, venv_root)
+        if cached_data:
+            if _is_stale_node_cache(cached_data):
+                pass
+            else:
+                try:
+                    flushed = await extension.flush_pending_routes()
+                    logger.info("][ %s flushed %d routes", extension_name, flushed)
+                except Exception as exc:
+                    logger.warning("][ %s route flush failed: %s", extension_name, exc)
+                specs: List[Tuple[str, str, type]] = []
+                for node_name, details in cached_data.items():
+                    stub_cls = build_stub_class(node_name, details, extension)
+                    specs.append(
+                        (node_name, details.get("display_name", node_name), stub_cls)
+                    )
+                return specs
+    # Cache miss - spawn process and get metadata
+
+    try:
+        remote_nodes: Dict[str, str] = await extension.list_nodes()
+    except Exception as exc:
+        logger.warning(
+            "][ %s metadata discovery failed, skipping isolated load: %s",
+            extension_name,
+            exc,
+        )
+        await _stop_extension_safe(extension, extension_name)
+        return []
+
+    if not remote_nodes:
+        logger.debug("][ %s exposed no isolated nodes; skipping", extension_name)
+        await _stop_extension_safe(extension, extension_name)
+        return []
+
+    specs: List[Tuple[str, str, type]] = []
+    cache_data: Dict[str, Dict] = {}
+
+    for node_name, display_name in remote_nodes.items():
+        try:
+            details = await extension.get_node_details(node_name)
+        except Exception as exc:
+            logger.warning(
+                "][ %s failed to load metadata for %s, skipping node: %s",
+                extension_name,
+                node_name,
+                exc,
+            )
+            continue
+        details["display_name"] = display_name
+        cache_data[node_name] = details
+        stub_cls = build_stub_class(node_name, details, extension)
+        specs.append((node_name, display_name, stub_cls))
+
+    if not specs:
+        logger.warning(
+            "][ %s produced no usable nodes after metadata scan; skipping",
+            extension_name,
+        )
+        await _stop_extension_safe(extension, extension_name)
+        return []
+
+    # Save metadata to cache for future runs
+    save_to_cache(node_dir, venv_root, cache_data, manifest_path)
+    logger.debug(f"][ {extension_name} metadata cached")
+
+    # Re-check web directory AFTER child has populated it
+    if host_policy["sandbox_mode"] == "disabled":
+        _register_web_directory(extension_name, node_dir)
+
+    # Flush any routes the child buffered during module import — must happen
+    # before router freeze and before we kill the child process.
+    try:
+        flushed = await extension.flush_pending_routes()
+        logger.info("][ %s flushed %d routes", extension_name, flushed)
+    except Exception as exc:
+        logger.warning("][ %s route flush failed: %s", extension_name, exc)
+
+    # EJECT: Kill process after getting metadata (will respawn on first execution)
+    await _stop_extension_safe(extension, extension_name)
+
+    return specs
+
+
+__all__ = ["ExtensionLoadError", "register_dummy_module", "load_isolated_node"]

comfy/isolation/extension_wrapper.py

@@ -0,0 +1,942 @@
+# pylint: disable=consider-using-from-import,cyclic-import,import-outside-toplevel,logging-fstring-interpolation,protected-access,wrong-import-position
+from __future__ import annotations
+
+import asyncio
+import torch
+
+
+class AttrDict(dict):
+    def __getattr__(self, item):
+        try:
+            return self[item]
+        except KeyError as e:
+            raise AttributeError(item) from e
+
+    def copy(self):
+        return AttrDict(super().copy())
+
+
+import importlib
+import inspect
+import json
+import logging
+import os
+import sys
+import uuid
+from dataclasses import asdict
+from typing import Any, Dict, List, Tuple
+
+from pyisolate import ExtensionBase
+
+from comfy_api.internal import _ComfyNodeInternal
+
+LOG_PREFIX = "]["
+V3_DISCOVERY_TIMEOUT = 30
+_PRE_EXEC_MIN_FREE_VRAM_BYTES = 2 * 1024 * 1024 * 1024
+
+logger = logging.getLogger(__name__)
+
+
+def _run_prestartup_web_copy(module: Any, module_dir: str, web_dir_path: str) -> None:
+    """Run the web asset copy step that prestartup_script.py used to do.
+
+    If the module's web/ directory is empty and the module had a
+    prestartup_script.py that copied assets from pip packages, this
+    function replicates that work inside the child process.
+
+    Generic pattern: reads _PRESTARTUP_WEB_COPY from the module if
+    defined, otherwise falls back to detecting common asset packages.
+    """
+    import shutil
+
+    # Already populated — nothing to do
+    if os.path.isdir(web_dir_path) and any(os.scandir(web_dir_path)):
+        return
+
+    os.makedirs(web_dir_path, exist_ok=True)
+
+    # Try module-defined copy spec first (generic hook for any node pack)
+    copy_spec = getattr(module, "_PRESTARTUP_WEB_COPY", None)
+    if copy_spec is not None and callable(copy_spec):
+        try:
+            copy_spec(web_dir_path)
+            logger.info(
+                "%s Ran _PRESTARTUP_WEB_COPY for %s", LOG_PREFIX, module_dir
+            )
+            return
+        except Exception as e:
+            logger.warning(
+                "%s _PRESTARTUP_WEB_COPY failed for %s: %s",
+                LOG_PREFIX, module_dir, e,
+            )
+
+    # Fallback: detect comfy_3d_viewers and run copy_viewer()
+    try:
+        from comfy_3d_viewers import copy_viewer, VIEWER_FILES
+        viewers = list(VIEWER_FILES.keys())
+        for viewer in viewers:
+            try:
+                copy_viewer(viewer, web_dir_path)
+            except Exception:
+                pass
+        if any(os.scandir(web_dir_path)):
+            logger.info(
+                "%s Copied %d viewer types from comfy_3d_viewers to %s",
+                LOG_PREFIX, len(viewers), web_dir_path,
+            )
+    except ImportError:
+        pass
+
+    # Fallback: detect comfy_dynamic_widgets
+    try:
+        from comfy_dynamic_widgets import get_js_path
+        src = os.path.realpath(get_js_path())
+        if os.path.exists(src):
+            dst_dir = os.path.join(web_dir_path, "js")
+            os.makedirs(dst_dir, exist_ok=True)
+            dst = os.path.join(dst_dir, "dynamic_widgets.js")
+            shutil.copy2(src, dst)
+    except ImportError:
+        pass
+
+
+def _read_extension_name(module_dir: str) -> str:
+    """Read extension name from pyproject.toml, falling back to directory name."""
+    pyproject = os.path.join(module_dir, "pyproject.toml")
+    if os.path.exists(pyproject):
+        try:
+            import tomllib
+        except ImportError:
+            import tomli as tomllib  # type: ignore[no-redef]
+        try:
+            with open(pyproject, "rb") as f:
+                data = tomllib.load(f)
+            name = data.get("project", {}).get("name")
+            if name:
+                return name
+        except Exception:
+            pass
+    return os.path.basename(module_dir)
+
+
+def _flush_tensor_transport_state(marker: str) -> int:
+    try:
+        from pyisolate import flush_tensor_keeper  # type: ignore[attr-defined]
+    except Exception:
+        return 0
+    if not callable(flush_tensor_keeper):
+        return 0
+    flushed = flush_tensor_keeper()
+    if flushed > 0:
+        logger.debug(
+            "%s %s flush_tensor_keeper released=%d", LOG_PREFIX, marker, flushed
+        )
+    return flushed
+
+
+def _relieve_child_vram_pressure(marker: str) -> None:
+    import comfy.model_management as model_management
+
+    model_management.cleanup_models_gc()
+    model_management.cleanup_models()
+
+    device = model_management.get_torch_device()
+    if not hasattr(device, "type") or device.type == "cpu":
+        return
+
+    required = max(
+        model_management.minimum_inference_memory(),
+        _PRE_EXEC_MIN_FREE_VRAM_BYTES,
+    )
+    if model_management.get_free_memory(device) < required:
+        model_management.free_memory(required, device, for_dynamic=True)
+        if model_management.get_free_memory(device) < required:
+            model_management.free_memory(required, device, for_dynamic=False)
+        model_management.cleanup_models()
+        model_management.soft_empty_cache()
+        logger.debug("%s %s free_memory target=%d", LOG_PREFIX, marker, required)
+
+
+def _sanitize_for_transport(value):
+    primitives = (str, int, float, bool, type(None))
+    if isinstance(value, primitives):
+        return value
+
+    cls_name = value.__class__.__name__
+    if cls_name == "FlexibleOptionalInputType":
+        return {
+            "__pyisolate_flexible_optional__": True,
+            "type": _sanitize_for_transport(getattr(value, "type", "*")),
+        }
+    if cls_name == "AnyType":
+        return {"__pyisolate_any_type__": True, "value": str(value)}
+    if cls_name == "ByPassTypeTuple":
+        return {
+            "__pyisolate_bypass_tuple__": [
+                _sanitize_for_transport(v) for v in tuple(value)
+            ]
+        }
+
+    if isinstance(value, dict):
+        return {k: _sanitize_for_transport(v) for k, v in value.items()}
+    if isinstance(value, tuple):
+        return {"__pyisolate_tuple__": [_sanitize_for_transport(v) for v in value]}
+    if isinstance(value, list):
+        return [_sanitize_for_transport(v) for v in value]
+
+    return str(value)
+
+
+# Re-export RemoteObjectHandle from pyisolate for backward compatibility
+# The canonical definition is now in pyisolate._internal.remote_handle
+from pyisolate._internal.remote_handle import RemoteObjectHandle  # noqa: E402,F401
+
+
+class ComfyNodeExtension(ExtensionBase):
+    def __init__(self) -> None:
+        super().__init__()
+        self.node_classes: Dict[str, type] = {}
+        self.display_names: Dict[str, str] = {}
+        self.node_instances: Dict[str, Any] = {}
+        self.remote_objects: Dict[str, Any] = {}
+        self._route_handlers: Dict[str, Any] = {}
+        self._module: Any = None
+        self._metadata_ready = asyncio.Event()
+
+    async def on_module_loaded(self, module: Any) -> None:
+        try:
+            self._module = module
+
+            # Registries are initialized in host_hooks.py initialize_host_process()
+            # They auto-register via ProxiedSingleton when instantiated
+            # NO additional setup required here - if a registry is missing from host_hooks, it WILL fail
+
+            self.node_classes = getattr(module, "NODE_CLASS_MAPPINGS", {}) or {}
+            self.display_names = getattr(module, "NODE_DISPLAY_NAME_MAPPINGS", {}) or {}
+            self._register_module_routes(module)
+
+            # Register web directory with WebDirectoryProxy (child-side)
+            web_dir_attr = getattr(module, "WEB_DIRECTORY", None)
+            if web_dir_attr is not None:
+                module_dir = os.path.dirname(os.path.abspath(module.__file__))
+                web_dir_path = os.path.abspath(os.path.join(module_dir, web_dir_attr))
+                ext_name = _read_extension_name(module_dir)
+
+                # If web dir is empty, run the copy step that prestartup_script.py did
+                _run_prestartup_web_copy(module, module_dir, web_dir_path)
+
+                if os.path.isdir(web_dir_path) and any(os.scandir(web_dir_path)):
+                    from comfy.isolation.proxies.web_directory_proxy import WebDirectoryProxy
+                    WebDirectoryProxy.register_web_dir(ext_name, web_dir_path)
+
+            try:
+                from comfy_api.latest import ComfyExtension
+
+                for name, obj in inspect.getmembers(module):
+                    if not (
+                        inspect.isclass(obj)
+                        and issubclass(obj, ComfyExtension)
+                        and obj is not ComfyExtension
+                    ):
+                        continue
+                    if not obj.__module__.startswith(module.__name__):
+                        continue
+                    try:
+                        ext_instance = obj()
+                        try:
+                            await asyncio.wait_for(
+                                ext_instance.on_load(), timeout=V3_DISCOVERY_TIMEOUT
+                            )
+                        except asyncio.TimeoutError:
+                            logger.error(
+                                "%s V3 Extension %s timed out in on_load()",
+                                LOG_PREFIX,
+                                name,
+                            )
+                            continue
+                        try:
+                            v3_nodes = await asyncio.wait_for(
+                                ext_instance.get_node_list(), timeout=V3_DISCOVERY_TIMEOUT
+                            )
+                        except asyncio.TimeoutError:
+                            logger.error(
+                                "%s V3 Extension %s timed out in get_node_list()",
+                                LOG_PREFIX,
+                                name,
+                            )
+                            continue
+                        for node_cls in v3_nodes:
+                            if hasattr(node_cls, "GET_SCHEMA"):
+                                schema = node_cls.GET_SCHEMA()
+                                self.node_classes[schema.node_id] = node_cls
+                                if schema.display_name:
+                                    self.display_names[schema.node_id] = schema.display_name
+                    except Exception as e:
+                        logger.error("%s V3 Extension %s failed: %s", LOG_PREFIX, name, e)
+            except ImportError:
+                pass
+
+            module_name = getattr(module, "__name__", "isolated_nodes")
+            for node_cls in self.node_classes.values():
+                if hasattr(node_cls, "__module__") and "/" in str(node_cls.__module__):
+                    node_cls.__module__ = module_name
+
+            self.node_instances = {}
+        finally:
+            self._metadata_ready.set()
+
+    def _register_module_routes(self, module: Any) -> None:
+        """Bridge legacy module-level ROUTES declarations into isolated routing."""
+        routes = getattr(module, "ROUTES", None) or []
+        if not routes:
+            return
+
+        from comfy.isolation.proxies.prompt_server_impl import PromptServerStub
+
+        prompt_server = PromptServerStub()
+        route_table = getattr(prompt_server, "routes", None)
+        if route_table is None:
+            logger.warning("%s Route registration unavailable for %s", LOG_PREFIX, module)
+            return
+
+        for route_spec in routes:
+            if not isinstance(route_spec, dict):
+                logger.warning("%s Ignoring non-dict ROUTES entry: %r", LOG_PREFIX, route_spec)
+                continue
+
+            method = str(route_spec.get("method", "")).strip().upper()
+            path = str(route_spec.get("path", "")).strip()
+            handler_ref = route_spec.get("handler")
+            if not method or not path:
+                logger.warning("%s Ignoring incomplete route spec: %r", LOG_PREFIX, route_spec)
+                continue
+
+            if isinstance(handler_ref, str):
+                handler = getattr(module, handler_ref, None)
+            else:
+                handler = handler_ref
+            if not callable(handler):
+                logger.warning(
+                    "%s Ignoring route with missing handler %r for %s %s",
+                    LOG_PREFIX,
+                    handler_ref,
+                    method,
+                    path,
+                )
+                continue
+
+            decorator = getattr(route_table, method.lower(), None)
+            if not callable(decorator):
+                logger.warning("%s Unsupported route method %s for %s", LOG_PREFIX, method, path)
+                continue
+
+            decorator(path)(handler)
+            self._route_handlers[f"{method} {path}"] = handler
+            logger.info("%s buffered legacy route %s %s", LOG_PREFIX, method, path)
+
+    async def list_nodes(self) -> Dict[str, str]:
+        await asyncio.wait_for(
+            self._metadata_ready.wait(), timeout=V3_DISCOVERY_TIMEOUT
+        )
+        return {name: self.display_names.get(name, name) for name in self.node_classes}
+
+    async def get_node_info(self, node_name: str) -> Dict[str, Any]:
+        return await self.get_node_details(node_name)
+
+    async def get_node_details(self, node_name: str) -> Dict[str, Any]:
+        await asyncio.wait_for(
+            self._metadata_ready.wait(), timeout=V3_DISCOVERY_TIMEOUT
+        )
+        node_cls = self._get_node_class(node_name)
+        is_v3 = issubclass(node_cls, _ComfyNodeInternal)
+
+        input_types_raw = (
+            node_cls.INPUT_TYPES() if hasattr(node_cls, "INPUT_TYPES") else {}
+        )
+        output_is_list = getattr(node_cls, "OUTPUT_IS_LIST", None)
+        if output_is_list is not None:
+            output_is_list = tuple(bool(x) for x in output_is_list)
+
+        details: Dict[str, Any] = {
+            "input_types": _sanitize_for_transport(input_types_raw),
+            "return_types": tuple(
+                str(t) for t in getattr(node_cls, "RETURN_TYPES", ())
+            ),
+            "return_names": getattr(node_cls, "RETURN_NAMES", None),
+            "function": str(getattr(node_cls, "FUNCTION", "execute")),
+            "category": str(getattr(node_cls, "CATEGORY", "")),
+            "output_node": bool(getattr(node_cls, "OUTPUT_NODE", False)),
+            "output_is_list": output_is_list,
+            "is_v3": is_v3,
+        }
+
+        if is_v3:
+            try:
+                schema = node_cls.GET_SCHEMA()
+                schema_v1 = asdict(schema.get_v1_info(node_cls))
+                try:
+                    schema_v3 = asdict(schema.get_v3_info(node_cls))
+                except (AttributeError, TypeError):
+                    schema_v3 = self._build_schema_v3_fallback(schema)
+                details.update(
+                    {
+                        "schema_v1": schema_v1,
+                        "schema_v3": schema_v3,
+                        "hidden": [h.value for h in (schema.hidden or [])],
+                        "description": getattr(schema, "description", ""),
+                        "deprecated": bool(getattr(node_cls, "DEPRECATED", False)),
+                        "experimental": bool(getattr(node_cls, "EXPERIMENTAL", False)),
+                        "api_node": bool(getattr(node_cls, "API_NODE", False)),
+                        "input_is_list": bool(
+                            getattr(node_cls, "INPUT_IS_LIST", False)
+                        ),
+                        "not_idempotent": bool(
+                            getattr(node_cls, "NOT_IDEMPOTENT", False)
+                        ),
+                        "accept_all_inputs": bool(
+                            getattr(node_cls, "ACCEPT_ALL_INPUTS", False)
+                        ),
+                    }
+                )
+            except Exception as exc:
+                logger.warning(
+                    "%s V3 schema serialization failed for %s: %s",
+                    LOG_PREFIX,
+                    node_name,
+                    exc,
+                )
+        return details
+
+    def _build_schema_v3_fallback(self, schema) -> Dict[str, Any]:
+        input_dict: Dict[str, Any] = {}
+        output_dict: Dict[str, Any] = {}
+        hidden_list: List[str] = []
+
+        if getattr(schema, "inputs", None):
+            for inp in schema.inputs:
+                self._add_schema_io_v3(inp, input_dict)
+        if getattr(schema, "outputs", None):
+            for out in schema.outputs:
+                self._add_schema_io_v3(out, output_dict)
+        if getattr(schema, "hidden", None):
+            for h in schema.hidden:
+                hidden_list.append(getattr(h, "value", str(h)))
+
+        return {
+            "input": input_dict,
+            "output": output_dict,
+            "hidden": hidden_list,
+            "name": getattr(schema, "node_id", None),
+            "display_name": getattr(schema, "display_name", None),
+            "description": getattr(schema, "description", None),
+            "category": getattr(schema, "category", None),
+            "output_node": getattr(schema, "is_output_node", False),
+            "deprecated": getattr(schema, "is_deprecated", False),
+            "experimental": getattr(schema, "is_experimental", False),
+            "api_node": getattr(schema, "is_api_node", False),
+        }
+
+    def _add_schema_io_v3(self, io_obj: Any, target: Dict[str, Any]) -> None:
+        io_id = getattr(io_obj, "id", None)
+        if io_id is None:
+            return
+
+        io_type_fn = getattr(io_obj, "get_io_type", None)
+        io_type = (
+            io_type_fn() if callable(io_type_fn) else getattr(io_obj, "io_type", None)
+        )
+
+        as_dict_fn = getattr(io_obj, "as_dict", None)
+        payload = as_dict_fn() if callable(as_dict_fn) else {}
+
+        target[str(io_id)] = (io_type, payload)
+
+    async def get_input_types(self, node_name: str) -> Dict[str, Any]:
+        node_cls = self._get_node_class(node_name)
+        if hasattr(node_cls, "INPUT_TYPES"):
+            return node_cls.INPUT_TYPES()
+        return {}
+
+    async def execute_node(self, node_name: str, **inputs: Any) -> Tuple[Any, ...]:
+        logger.debug(
+            "%s ISO:child_execute_start ext=%s node=%s input_keys=%d",
+            LOG_PREFIX,
+            getattr(self, "name", "?"),
+            node_name,
+            len(inputs),
+        )
+        if os.environ.get("PYISOLATE_CHILD") == "1":
+            _relieve_child_vram_pressure("EXT:pre_execute")
+
+        resolved_inputs = self._resolve_remote_objects(inputs)
+
+        instance = self._get_node_instance(node_name)
+        node_cls = self._get_node_class(node_name)
+
+        # V3 API nodes expect hidden parameters in cls.hidden, not as kwargs
+        # Hidden params come through RPC as string keys like "Hidden.prompt"
+        from comfy_api.latest._io import Hidden, HiddenHolder
+
+        # Map string representations back to Hidden enum keys
+        hidden_string_map = {
+            "Hidden.unique_id": Hidden.unique_id,
+            "Hidden.prompt": Hidden.prompt,
+            "Hidden.extra_pnginfo": Hidden.extra_pnginfo,
+            "Hidden.dynprompt": Hidden.dynprompt,
+            "Hidden.auth_token_comfy_org": Hidden.auth_token_comfy_org,
+            "Hidden.api_key_comfy_org": Hidden.api_key_comfy_org,
+            # Uppercase enum VALUE forms — V3 execution engine passes these
+            "UNIQUE_ID": Hidden.unique_id,
+            "PROMPT": Hidden.prompt,
+            "EXTRA_PNGINFO": Hidden.extra_pnginfo,
+            "DYNPROMPT": Hidden.dynprompt,
+            "AUTH_TOKEN_COMFY_ORG": Hidden.auth_token_comfy_org,
+            "API_KEY_COMFY_ORG": Hidden.api_key_comfy_org,
+        }
+
+        # Find and extract hidden parameters (both enum and string form)
+        hidden_found = {}
+        keys_to_remove = []
+
+        for key in list(resolved_inputs.keys()):
+            # Check string form first (from RPC serialization)
+            if key in hidden_string_map:
+                hidden_found[hidden_string_map[key]] = resolved_inputs[key]
+                keys_to_remove.append(key)
+            # Also check enum form (direct calls)
+            elif isinstance(key, Hidden):
+                hidden_found[key] = resolved_inputs[key]
+                keys_to_remove.append(key)
+
+        # Remove hidden params from kwargs
+        for key in keys_to_remove:
+            resolved_inputs.pop(key)
+
+        # Set hidden on node class if any hidden params found
+        if hidden_found:
+            if not hasattr(node_cls, "hidden") or node_cls.hidden is None:
+                node_cls.hidden = HiddenHolder.from_dict(hidden_found)
+            else:
+                # Update existing hidden holder
+                for key, value in hidden_found.items():
+                    setattr(node_cls.hidden, key.value.lower(), value)
+
+        # INPUT_IS_LIST: ComfyUI's executor passes all inputs as lists when this
+        # flag is set.  The isolation RPC delivers unwrapped values, so we must
+        # wrap each input in a single-element list to match the contract.
+        if getattr(node_cls, "INPUT_IS_LIST", False):
+            resolved_inputs = {k: [v] for k, v in resolved_inputs.items()}
+
+        function_name = getattr(node_cls, "FUNCTION", "execute")
+        if not hasattr(instance, function_name):
+            raise AttributeError(f"Node {node_name} missing callable '{function_name}'")
+
+        handler = getattr(instance, function_name)
+
+        try:
+            import torch
+            if asyncio.iscoroutinefunction(handler):
+                with torch.inference_mode():
+                    result = await handler(**resolved_inputs)
+            else:
+                import functools
+
+                def _run_with_inference_mode(**kwargs):
+                    with torch.inference_mode():
+                        return handler(**kwargs)
+
+                loop = asyncio.get_running_loop()
+                result = await loop.run_in_executor(
+                    None, functools.partial(_run_with_inference_mode, **resolved_inputs)
+                )
+        except Exception:
+            logger.exception(
+                "%s ISO:child_execute_error ext=%s node=%s",
+                LOG_PREFIX,
+                getattr(self, "name", "?"),
+                node_name,
+            )
+            raise
+
+        if type(result).__name__ == "NodeOutput":
+            node_output_dict = {
+                "__node_output__": True,
+                "args": self._wrap_unpicklable_objects(result.args),
+            }
+            if result.ui is not None:
+                node_output_dict["ui"] = self._wrap_unpicklable_objects(result.ui)
+            if getattr(result, "expand", None) is not None:
+                node_output_dict["expand"] = result.expand
+            if getattr(result, "block_execution", None) is not None:
+                node_output_dict["block_execution"] = result.block_execution
+            return node_output_dict
+        if self._is_comfy_protocol_return(result):
+            wrapped = self._wrap_unpicklable_objects(result)
+            return wrapped
+
+        if not isinstance(result, tuple):
+            result = (result,)
+        wrapped = self._wrap_unpicklable_objects(result)
+        return wrapped
+
+    async def flush_pending_routes(self) -> int:
+        """Flush buffered route registrations to host via RPC. Called by host after node discovery."""
+        from comfy.isolation.proxies.prompt_server_impl import PromptServerStub
+        return await PromptServerStub.flush_child_routes()
+
+    async def flush_transport_state(self) -> int:
+        if os.environ.get("PYISOLATE_CHILD") != "1":
+            return 0
+        logger.debug(
+            "%s ISO:child_flush_start ext=%s", LOG_PREFIX, getattr(self, "name", "?")
+        )
+        flushed = _flush_tensor_transport_state("EXT:workflow_end")
+        try:
+            from comfy.isolation.model_patcher_proxy_registry import (
+                ModelPatcherRegistry,
+            )
+
+            registry = ModelPatcherRegistry()
+            removed = registry.sweep_pending_cleanup()
+            if removed > 0:
+                logger.debug(
+                    "%s EXT:workflow_end registry sweep removed=%d", LOG_PREFIX, removed
+                )
+        except Exception:
+            logger.debug(
+                "%s EXT:workflow_end registry sweep failed", LOG_PREFIX, exc_info=True
+            )
+        logger.debug(
+            "%s ISO:child_flush_done ext=%s flushed=%d",
+            LOG_PREFIX,
+            getattr(self, "name", "?"),
+            flushed,
+        )
+        return flushed
+
+    async def get_remote_object(self, object_id: str) -> Any:
+        """Retrieve a remote object by ID for host-side deserialization."""
+        if object_id not in self.remote_objects:
+            raise KeyError(f"Remote object {object_id} not found")
+
+        return self.remote_objects[object_id]
+
+    def _store_remote_object_handle(self, obj: Any) -> RemoteObjectHandle:
+        object_id = str(uuid.uuid4())
+        self.remote_objects[object_id] = obj
+        return RemoteObjectHandle(object_id, type(obj).__name__)
+
+    async def call_remote_object_method(
+        self,
+        object_id: str,
+        method_name: str,
+        *args: Any,
+        **kwargs: Any,
+    ) -> Any:
+        """Invoke a method or attribute-backed accessor on a child-owned object."""
+        obj = await self.get_remote_object(object_id)
+
+        if method_name == "get_patcher_attr":
+            return getattr(obj, args[0])
+        if method_name == "get_model_options":
+            return getattr(obj, "model_options")
+        if method_name == "set_model_options":
+            setattr(obj, "model_options", args[0])
+            return None
+        if method_name == "get_object_patches":
+            return getattr(obj, "object_patches")
+        if method_name == "get_patches":
+            return getattr(obj, "patches")
+        if method_name == "get_wrappers":
+            return getattr(obj, "wrappers")
+        if method_name == "get_callbacks":
+            return getattr(obj, "callbacks")
+        if method_name == "get_load_device":
+            return getattr(obj, "load_device")
+        if method_name == "get_offload_device":
+            return getattr(obj, "offload_device")
+        if method_name == "get_hook_mode":
+            return getattr(obj, "hook_mode")
+        if method_name == "get_parent":
+            parent = getattr(obj, "parent", None)
+            if parent is None:
+                return None
+            return self._store_remote_object_handle(parent)
+        if method_name == "get_inner_model_attr":
+            attr_name = args[0]
+            if hasattr(obj.model, attr_name):
+                return getattr(obj.model, attr_name)
+            if hasattr(obj, attr_name):
+                return getattr(obj, attr_name)
+            return None
+        if method_name == "inner_model_apply_model":
+            return obj.model.apply_model(*args[0], **args[1])
+        if method_name == "inner_model_extra_conds_shapes":
+            return obj.model.extra_conds_shapes(*args[0], **args[1])
+        if method_name == "inner_model_extra_conds":
+            return obj.model.extra_conds(*args[0], **args[1])
+        if method_name == "inner_model_memory_required":
+            return obj.model.memory_required(*args[0], **args[1])
+        if method_name == "process_latent_in":
+            return obj.model.process_latent_in(*args[0], **args[1])
+        if method_name == "process_latent_out":
+            return obj.model.process_latent_out(*args[0], **args[1])
+        if method_name == "scale_latent_inpaint":
+            return obj.model.scale_latent_inpaint(*args[0], **args[1])
+        if method_name.startswith("get_"):
+            attr_name = method_name[4:]
+            if hasattr(obj, attr_name):
+                return getattr(obj, attr_name)
+
+        target = getattr(obj, method_name)
+        if callable(target):
+            result = target(*args, **kwargs)
+            if inspect.isawaitable(result):
+                result = await result
+            if type(result).__name__ == "ModelPatcher":
+                return self._store_remote_object_handle(result)
+            return result
+        if args or kwargs:
+            raise TypeError(f"{method_name} is not callable on remote object {object_id}")
+        return target
+
+    def _wrap_unpicklable_objects(self, data: Any) -> Any:
+        if isinstance(data, (str, int, float, bool, type(None))):
+            return data
+        if isinstance(data, torch.Tensor):
+            tensor = data.detach() if data.requires_grad else data
+            if os.environ.get("PYISOLATE_CHILD") == "1" and tensor.device.type != "cpu":
+                return tensor.cpu()
+            return tensor
+
+        # Special-case clip vision outputs: preserve attribute access by packing fields
+        if hasattr(data, "penultimate_hidden_states") or hasattr(
+            data, "last_hidden_state"
+        ):
+            fields = {}
+            for attr in (
+                "penultimate_hidden_states",
+                "last_hidden_state",
+                "image_embeds",
+                "text_embeds",
+            ):
+                if hasattr(data, attr):
+                    try:
+                        fields[attr] = self._wrap_unpicklable_objects(
+                            getattr(data, attr)
+                        )
+                    except Exception:
+                        pass
+            if fields:
+                return {"__pyisolate_attribute_container__": True, "data": fields}
+
+        # Avoid converting arbitrary objects with stateful methods (models, etc.)
+        # They will be handled via RemoteObjectHandle below.
+
+        type_name = type(data).__name__
+        if type_name == "ModelPatcherProxy":
+            return {"__type__": "ModelPatcherRef", "model_id": data._instance_id}
+        if type_name == "CLIPProxy":
+            return {"__type__": "CLIPRef", "clip_id": data._instance_id}
+        if type_name == "VAEProxy":
+            return {"__type__": "VAERef", "vae_id": data._instance_id}
+        if type_name == "ModelSamplingProxy":
+            return {"__type__": "ModelSamplingRef", "ms_id": data._instance_id}
+
+        if isinstance(data, (list, tuple)):
+            wrapped = [self._wrap_unpicklable_objects(item) for item in data]
+            return tuple(wrapped) if isinstance(data, tuple) else wrapped
+        if isinstance(data, dict):
+            converted_dict = {
+                k: self._wrap_unpicklable_objects(v) for k, v in data.items()
+            }
+            return {"__pyisolate_attrdict__": True, "data": converted_dict}
+
+        from pyisolate._internal.serialization_registry import SerializerRegistry
+
+        registry = SerializerRegistry.get_instance()
+        if registry.is_data_type(type_name):
+            serializer = registry.get_serializer(type_name)
+            if serializer:
+                return serializer(data)
+
+        return self._store_remote_object_handle(data)
+
+    def _resolve_remote_objects(self, data: Any) -> Any:
+        if isinstance(data, RemoteObjectHandle):
+            if data.object_id not in self.remote_objects:
+                raise KeyError(f"Remote object {data.object_id} not found")
+            return self.remote_objects[data.object_id]
+
+        if isinstance(data, dict):
+            ref_type = data.get("__type__")
+            if ref_type in ("CLIPRef", "ModelPatcherRef", "VAERef"):
+                from pyisolate._internal.model_serialization import (
+                    deserialize_proxy_result,
+                )
+
+                return deserialize_proxy_result(data)
+            if ref_type == "ModelSamplingRef":
+                from pyisolate._internal.model_serialization import (
+                    deserialize_proxy_result,
+                )
+
+                return deserialize_proxy_result(data)
+            return {k: self._resolve_remote_objects(v) for k, v in data.items()}
+
+        if isinstance(data, (list, tuple)):
+            resolved = [self._resolve_remote_objects(item) for item in data]
+            return tuple(resolved) if isinstance(data, tuple) else resolved
+        return data
+
+    def _get_node_class(self, node_name: str) -> type:
+        if node_name not in self.node_classes:
+            raise KeyError(f"Unknown node: {node_name}")
+        return self.node_classes[node_name]
+
+    def _get_node_instance(self, node_name: str) -> Any:
+        if node_name not in self.node_instances:
+            if node_name not in self.node_classes:
+                raise KeyError(f"Unknown node: {node_name}")
+            self.node_instances[node_name] = self.node_classes[node_name]()
+        return self.node_instances[node_name]
+
+    async def before_module_loaded(self) -> None:
+        try:
+            from comfy.isolation import initialize_proxies
+
+            initialize_proxies()
+        except Exception as e:
+            logger.error(
+                "%s before_module_loaded initialize_proxies FAILED: %s", LOG_PREFIX, e
+            )
+
+        await super().before_module_loaded()
+        try:
+            from comfy_api.latest import ComfyAPI_latest
+            from .proxies.progress_proxy import ProgressProxy
+
+            ComfyAPI_latest.Execution = ProgressProxy
+            # ComfyAPI_latest.execution = ProgressProxy()  # Eliminated to avoid Singleton collision
+            # fp_proxy = FolderPathsProxy()                 # Eliminated to avoid Singleton collision
+            # latest_ui.folder_paths = fp_proxy
+            # latest_resources.folder_paths = fp_proxy
+        except Exception:
+            pass
+
+    async def call_route_handler(
+        self,
+        handler_module: str,
+        handler_func: str,
+        request_data: Dict[str, Any],
+    ) -> Any:
+        cache_key = f"{handler_module}.{handler_func}"
+        if cache_key not in self._route_handlers:
+            if self._module is not None and hasattr(self._module, "__file__"):
+                node_dir = os.path.dirname(self._module.__file__)
+                if node_dir not in sys.path:
+                    sys.path.insert(0, node_dir)
+            try:
+                module = importlib.import_module(handler_module)
+                self._route_handlers[cache_key] = getattr(module, handler_func)
+            except (ImportError, AttributeError) as e:
+                raise ValueError(f"Route handler not found: {cache_key}") from e
+
+        handler = self._route_handlers[cache_key]
+        mock_request = MockRequest(request_data)
+
+        if asyncio.iscoroutinefunction(handler):
+            result = await handler(mock_request)
+        else:
+            result = handler(mock_request)
+        return self._serialize_response(result)
+
+    def _is_comfy_protocol_return(self, result: Any) -> bool:
+        """
+        Check if the result matches the ComfyUI 'Protocol Return' schema.
+
+        A Protocol Return is a dictionary containing specific reserved keys that
+        ComfyUI's execution engine interprets as instructions (UI updates,
+        Workflow expansion, etc.) rather than purely data outputs.
+
+        Schema:
+           - Must be a dict
+           - Must contain at least one of: 'ui', 'result', 'expand'
+        """
+        if not isinstance(result, dict):
+            return False
+        return any(key in result for key in ("ui", "result", "expand"))
+
+    def _serialize_response(self, response: Any) -> Dict[str, Any]:
+        if response is None:
+            return {"type": "text", "body": "", "status": 204}
+        if isinstance(response, dict):
+            return {"type": "json", "body": response, "status": 200}
+        if isinstance(response, str):
+            return {"type": "text", "body": response, "status": 200}
+        if hasattr(response, "text") and hasattr(response, "status"):
+            return {
+                "type": "text",
+                "body": response.text
+                if hasattr(response, "text")
+                else str(response.body),
+                "status": response.status,
+                "headers": dict(response.headers)
+                if hasattr(response, "headers")
+                else {},
+            }
+        if hasattr(response, "body") and hasattr(response, "status"):
+            body = response.body
+            if isinstance(body, bytes):
+                try:
+                    return {
+                        "type": "text",
+                        "body": body.decode("utf-8"),
+                        "status": response.status,
+                    }
+                except UnicodeDecodeError:
+                    return {
+                        "type": "binary",
+                        "body": body.hex(),
+                        "status": response.status,
+                    }
+            return {"type": "json", "body": body, "status": response.status}
+        return {"type": "text", "body": str(response), "status": 200}
+
+
+class MockRequest:
+    def __init__(self, data: Dict[str, Any]):
+        self.method = data.get("method", "GET")
+        self.path = data.get("path", "/")
+        self.query = data.get("query", {})
+        self._body = data.get("body", {})
+        self._text = data.get("text", "")
+        self.headers = data.get("headers", {})
+        self.content_type = data.get(
+            "content_type", self.headers.get("Content-Type", "application/json")
+        )
+        self.match_info = data.get("match_info", {})
+
+    async def json(self) -> Any:
+        if isinstance(self._body, dict):
+            return self._body
+        if isinstance(self._body, str):
+            return json.loads(self._body)
+        return {}
+
+    async def post(self) -> Dict[str, Any]:
+        if isinstance(self._body, dict):
+            return self._body
+        return {}
+
+    async def text(self) -> str:
+        if self._text:
+            return self._text
+        if isinstance(self._body, str):
+            return self._body
+        if isinstance(self._body, dict):
+            return json.dumps(self._body)
+        return ""
+
+    async def read(self) -> bytes:
+        return (await self.text()).encode("utf-8")

comfy/isolation/host_hooks.py

@@ -0,0 +1,25 @@
+# pylint: disable=import-outside-toplevel
+# Host process initialization for PyIsolate
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+def initialize_host_process() -> None:
+    from .proxies.folder_paths_proxy import FolderPathsProxy
+    from .proxies.helper_proxies import HelperProxiesService
+    from .proxies.model_management_proxy import ModelManagementProxy
+    from .proxies.progress_proxy import ProgressProxy
+    from .proxies.prompt_server_impl import PromptServerService
+    from .proxies.utils_proxy import UtilsProxy
+    from .proxies.web_directory_proxy import WebDirectoryProxy
+    from .vae_proxy import VAERegistry
+
+    FolderPathsProxy()
+    HelperProxiesService()
+    ModelManagementProxy()
+    ProgressProxy()
+    PromptServerService()
+    UtilsProxy()
+    WebDirectoryProxy()
+    VAERegistry()

comfy/isolation/host_policy.py

@@ -0,0 +1,180 @@
+# pylint: disable=logging-fstring-interpolation
+from __future__ import annotations
+
+import logging
+import os
+from pathlib import Path
+from pathlib import PurePosixPath
+from typing import Dict, List, TypedDict
+
+try:
+    import tomllib
+except ImportError:
+    import tomli as tomllib  # type: ignore[no-redef]
+
+logger = logging.getLogger(__name__)
+
+HOST_POLICY_PATH_ENV = "COMFY_HOST_POLICY_PATH"
+VALID_SANDBOX_MODES = frozenset({"required", "disabled"})
+FORBIDDEN_WRITABLE_PATHS = frozenset({"/tmp"})
+
+
+class HostSecurityPolicy(TypedDict):
+    sandbox_mode: str
+    allow_network: bool
+    writable_paths: List[str]
+    readonly_paths: List[str]
+    sealed_worker_ro_import_paths: List[str]
+    whitelist: Dict[str, str]
+
+
+DEFAULT_POLICY: HostSecurityPolicy = {
+    "sandbox_mode": "required",
+    "allow_network": False,
+    "writable_paths": ["/dev/shm"],
+    "readonly_paths": [],
+    "sealed_worker_ro_import_paths": [],
+    "whitelist": {},
+}
+
+
+def _default_policy() -> HostSecurityPolicy:
+    return {
+        "sandbox_mode": DEFAULT_POLICY["sandbox_mode"],
+        "allow_network": DEFAULT_POLICY["allow_network"],
+        "writable_paths": list(DEFAULT_POLICY["writable_paths"]),
+        "readonly_paths": list(DEFAULT_POLICY["readonly_paths"]),
+        "sealed_worker_ro_import_paths": list(DEFAULT_POLICY["sealed_worker_ro_import_paths"]),
+        "whitelist": dict(DEFAULT_POLICY["whitelist"]),
+    }
+
+
+def _normalize_writable_paths(paths: list[object]) -> list[str]:
+    normalized_paths: list[str] = []
+    for raw_path in paths:
+        # Host-policy paths are contract-style POSIX paths; keep representation
+        # stable across Windows/Linux so tests and config behavior stay consistent.
+        normalized_path = str(PurePosixPath(str(raw_path).replace("\\", "/")))
+        if normalized_path in FORBIDDEN_WRITABLE_PATHS:
+            continue
+        normalized_paths.append(normalized_path)
+    return normalized_paths
+
+
+def _load_whitelist_file(file_path: Path, config_path: Path) -> Dict[str, str]:
+    if not file_path.is_absolute():
+        file_path = config_path.parent / file_path
+    if not file_path.exists():
+        logger.warning("whitelist_file %s not found, skipping.", file_path)
+        return {}
+    entries: Dict[str, str] = {}
+    for line in file_path.read_text().splitlines():
+        line = line.strip()
+        if not line or line.startswith("#"):
+            continue
+        entries[line] = "*"
+    logger.debug("Loaded %d whitelist entries from %s", len(entries), file_path)
+    return entries
+
+
+def _normalize_sealed_worker_ro_import_paths(raw_paths: object) -> list[str]:
+    if not isinstance(raw_paths, list):
+        raise ValueError(
+            "tool.comfy.host.sealed_worker_ro_import_paths must be a list of absolute paths."
+        )
+
+    normalized_paths: list[str] = []
+    seen: set[str] = set()
+    for raw_path in raw_paths:
+        if not isinstance(raw_path, str) or not raw_path.strip():
+            raise ValueError(
+                "tool.comfy.host.sealed_worker_ro_import_paths entries must be non-empty strings."
+            )
+        normalized_path = str(PurePosixPath(raw_path.replace("\\", "/")))
+        # Accept both POSIX absolute paths (/home/...) and Windows drive-letter paths (D:/...)
+        is_absolute = normalized_path.startswith("/") or (
+            len(normalized_path) >= 3 and normalized_path[1] == ":" and normalized_path[2] == "/"
+        )
+        if not is_absolute:
+            raise ValueError(
+                "tool.comfy.host.sealed_worker_ro_import_paths entries must be absolute paths."
+            )
+        if normalized_path not in seen:
+            seen.add(normalized_path)
+            normalized_paths.append(normalized_path)
+
+    return normalized_paths
+
+
+def load_host_policy(comfy_root: Path) -> HostSecurityPolicy:
+    config_override = os.environ.get(HOST_POLICY_PATH_ENV)
+    config_path = Path(config_override) if config_override else comfy_root / "pyproject.toml"
+    policy = _default_policy()
+
+    if not config_path.exists():
+        logger.debug("Host policy file missing at %s, using defaults.", config_path)
+        return policy
+
+    try:
+        with config_path.open("rb") as f:
+            data = tomllib.load(f)
+    except Exception:
+        logger.warning(
+            "Failed to parse host policy from %s, using defaults.",
+            config_path,
+            exc_info=True,
+        )
+        return policy
+
+    tool_config = data.get("tool", {}).get("comfy", {}).get("host", {})
+    if not isinstance(tool_config, dict):
+        logger.debug("No [tool.comfy.host] section found, using defaults.")
+        return policy
+
+    sandbox_mode = tool_config.get("sandbox_mode")
+    if isinstance(sandbox_mode, str):
+        normalized_sandbox_mode = sandbox_mode.strip().lower()
+        if normalized_sandbox_mode in VALID_SANDBOX_MODES:
+            policy["sandbox_mode"] = normalized_sandbox_mode
+        else:
+            logger.warning(
+                "Invalid host sandbox_mode %r in %s, using default %r.",
+                sandbox_mode,
+                config_path,
+                DEFAULT_POLICY["sandbox_mode"],
+            )
+
+    if "allow_network" in tool_config:
+        policy["allow_network"] = bool(tool_config["allow_network"])
+
+    if "writable_paths" in tool_config:
+        policy["writable_paths"] = _normalize_writable_paths(tool_config["writable_paths"])
+
+    if "readonly_paths" in tool_config:
+        policy["readonly_paths"] = [str(p) for p in tool_config["readonly_paths"]]
+
+    if "sealed_worker_ro_import_paths" in tool_config:
+        policy["sealed_worker_ro_import_paths"] = _normalize_sealed_worker_ro_import_paths(
+            tool_config["sealed_worker_ro_import_paths"]
+        )
+
+    whitelist_file = tool_config.get("whitelist_file")
+    if isinstance(whitelist_file, str):
+        policy["whitelist"].update(_load_whitelist_file(Path(whitelist_file), config_path))
+
+    whitelist_raw = tool_config.get("whitelist")
+    if isinstance(whitelist_raw, dict):
+        policy["whitelist"].update({str(k): str(v) for k, v in whitelist_raw.items()})
+
+    os.environ["PYISOLATE_SANDBOX_MODE"] = policy["sandbox_mode"]
+
+    logger.debug(
+        "Loaded Host Policy: %d whitelisted nodes, Sandbox=%s, Network=%s",
+        len(policy["whitelist"]),
+        policy["sandbox_mode"],
+        policy["allow_network"],
+    )
+    return policy
+
+
+__all__ = ["HostSecurityPolicy", "load_host_policy", "DEFAULT_POLICY"]

comfy/isolation/manifest_loader.py

@@ -0,0 +1,221 @@
+# pylint: disable=import-outside-toplevel
+from __future__ import annotations
+
+import hashlib
+import json
+import logging
+import os
+import sys
+import tempfile
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Tuple
+
+import folder_paths
+
+try:
+    import tomllib
+except ImportError:
+    import tomli as tomllib  # type: ignore[no-redef]
+
+LOG_PREFIX = "]["
+logger = logging.getLogger(__name__)
+
+CACHE_SUBDIR = "cache"
+CACHE_KEY_FILE = "cache_key"
+CACHE_DATA_FILE = "node_info.json"
+CACHE_KEY_LENGTH = 16
+_NESTED_SCAN_ROOT = "packages"
+_IGNORED_MANIFEST_DIRS = {".git", ".venv", "__pycache__"}
+
+
+def _read_manifest(manifest_path: Path) -> dict[str, Any] | None:
+    try:
+        with manifest_path.open("rb") as f:
+            data = tomllib.load(f)
+        if isinstance(data, dict):
+            return data
+    except Exception:
+        return None
+    return None
+
+
+def _is_isolation_manifest(data: dict[str, Any]) -> bool:
+    return (
+        "tool" in data
+        and "comfy" in data["tool"]
+        and "isolation" in data["tool"]["comfy"]
+    )
+
+
+def _discover_nested_manifests(entry: Path) -> List[Tuple[Path, Path]]:
+    packages_root = entry / _NESTED_SCAN_ROOT
+    if not packages_root.exists() or not packages_root.is_dir():
+        return []
+
+    nested: List[Tuple[Path, Path]] = []
+    for manifest in sorted(packages_root.rglob("pyproject.toml")):
+        node_dir = manifest.parent
+        if any(part in _IGNORED_MANIFEST_DIRS for part in node_dir.parts):
+            continue
+
+        data = _read_manifest(manifest)
+        if not data or not _is_isolation_manifest(data):
+            continue
+
+        isolation = data["tool"]["comfy"]["isolation"]
+        if isolation.get("standalone") is True:
+            nested.append((node_dir, manifest))
+
+    return nested
+
+
+def find_manifest_directories() -> List[Tuple[Path, Path]]:
+    """Find custom node directories containing a valid pyproject.toml with [tool.comfy.isolation]."""
+    manifest_dirs: List[Tuple[Path, Path]] = []
+
+    # Standard custom_nodes paths
+    for base_path in folder_paths.get_folder_paths("custom_nodes"):
+        base = Path(base_path)
+        if not base.exists() or not base.is_dir():
+            continue
+
+        for entry in base.iterdir():
+            if not entry.is_dir():
+                continue
+
+            # Look for pyproject.toml
+            manifest = entry / "pyproject.toml"
+            if not manifest.exists():
+                continue
+
+            data = _read_manifest(manifest)
+            if not data or not _is_isolation_manifest(data):
+                continue
+
+            manifest_dirs.append((entry, manifest))
+            manifest_dirs.extend(_discover_nested_manifests(entry))
+
+    return manifest_dirs
+
+
+def compute_cache_key(node_dir: Path, manifest_path: Path) -> str:
+    """Hash manifest + .py mtimes + Python version + PyIsolate version."""
+    hasher = hashlib.sha256()
+
+    try:
+        # Hashing the manifest content ensures config changes invalidate cache
+        hasher.update(manifest_path.read_bytes())
+    except OSError:
+        hasher.update(b"__manifest_read_error__")
+
+    try:
+        py_files = sorted(node_dir.rglob("*.py"))
+        for py_file in py_files:
+            rel_path = py_file.relative_to(node_dir)
+            if "__pycache__" in str(rel_path) or ".venv" in str(rel_path):
+                continue
+            hasher.update(str(rel_path).encode("utf-8"))
+            try:
+                hasher.update(str(py_file.stat().st_mtime).encode("utf-8"))
+            except OSError:
+                hasher.update(b"__file_stat_error__")
+    except OSError:
+        hasher.update(b"__dir_scan_error__")
+
+    hasher.update(sys.version.encode("utf-8"))
+
+    try:
+        import pyisolate
+
+        hasher.update(pyisolate.__version__.encode("utf-8"))
+    except (ImportError, AttributeError):
+        hasher.update(b"__pyisolate_unknown__")
+
+    return hasher.hexdigest()[:CACHE_KEY_LENGTH]
+
+
+def get_cache_path(node_dir: Path, venv_root: Path) -> Tuple[Path, Path]:
+    """Return (cache_key_file, cache_data_file) in venv_root/{node}/cache/."""
+    cache_dir = venv_root / node_dir.name / CACHE_SUBDIR
+    return (cache_dir / CACHE_KEY_FILE, cache_dir / CACHE_DATA_FILE)
+
+
+def is_cache_valid(node_dir: Path, manifest_path: Path, venv_root: Path) -> bool:
+    """Return True only if stored cache key matches current computed key."""
+    try:
+        cache_key_file, cache_data_file = get_cache_path(node_dir, venv_root)
+        if not cache_key_file.exists() or not cache_data_file.exists():
+            return False
+        current_key = compute_cache_key(node_dir, manifest_path)
+        stored_key = cache_key_file.read_text(encoding="utf-8").strip()
+        return current_key == stored_key
+    except Exception as e:
+        logger.debug(
+            "%s Cache validation error for %s: %s", LOG_PREFIX, node_dir.name, e
+        )
+        return False
+
+
+def load_from_cache(node_dir: Path, venv_root: Path) -> Optional[Dict[str, Any]]:
+    """Load node metadata from cache, return None on any error."""
+    try:
+        _, cache_data_file = get_cache_path(node_dir, venv_root)
+        if not cache_data_file.exists():
+            return None
+        data = json.loads(cache_data_file.read_text(encoding="utf-8"))
+        if not isinstance(data, dict):
+            return None
+        return data
+    except Exception:
+        return None
+
+
+def save_to_cache(
+    node_dir: Path, venv_root: Path, node_data: Dict[str, Any], manifest_path: Path
+) -> None:
+    """Save node metadata and cache key atomically."""
+    try:
+        cache_key_file, cache_data_file = get_cache_path(node_dir, venv_root)
+        cache_dir = cache_key_file.parent
+        cache_dir.mkdir(parents=True, exist_ok=True)
+        cache_key = compute_cache_key(node_dir, manifest_path)
+
+        # Atomic write: data
+        tmp_data_fd, tmp_data_path = tempfile.mkstemp(dir=str(cache_dir), suffix=".tmp")
+        try:
+            with os.fdopen(tmp_data_fd, "w", encoding="utf-8") as f:
+                json.dump(node_data, f, indent=2)
+            os.replace(tmp_data_path, cache_data_file)
+        except Exception:
+            try:
+                os.unlink(tmp_data_path)
+            except OSError:
+                pass
+            raise
+
+        # Atomic write: key
+        tmp_key_fd, tmp_key_path = tempfile.mkstemp(dir=str(cache_dir), suffix=".tmp")
+        try:
+            with os.fdopen(tmp_key_fd, "w", encoding="utf-8") as f:
+                f.write(cache_key)
+            os.replace(tmp_key_path, cache_key_file)
+        except Exception:
+            try:
+                os.unlink(tmp_key_path)
+            except OSError:
+                pass
+            raise
+
+    except Exception as e:
+        logger.warning("%s Cache save failed for %s: %s", LOG_PREFIX, node_dir.name, e)
+
+
+__all__ = [
+    "LOG_PREFIX",
+    "find_manifest_directories",
+    "compute_cache_key",
+    "get_cache_path",
+    "is_cache_valid",
+    "load_from_cache",
+    "save_to_cache",
+]

comfy/isolation/model_patcher_proxy.py

@@ -0,0 +1,891 @@
+# pylint: disable=bare-except,consider-using-from-import,import-outside-toplevel,protected-access
+# RPC proxy for ModelPatcher (parent process)
+from __future__ import annotations
+
+import logging
+from typing import Any, Optional, List, Set, Dict, Callable
+
+from comfy.isolation.proxies.base import (
+    IS_CHILD_PROCESS,
+    BaseProxy,
+)
+from comfy.isolation.model_patcher_proxy_registry import (
+    ModelPatcherRegistry,
+    AutoPatcherEjector,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class ModelPatcherProxy(BaseProxy[ModelPatcherRegistry]):
+    _registry_class = ModelPatcherRegistry
+    __module__ = "comfy.model_patcher"
+    _APPLY_MODEL_GUARD_PADDING_BYTES = 32 * 1024 * 1024
+
+    def _spawn_related_proxy(self, instance_id: str) -> "ModelPatcherProxy":
+        proxy = ModelPatcherProxy(
+            instance_id,
+            self._registry,
+            manage_lifecycle=not IS_CHILD_PROCESS,
+        )
+        if getattr(self, "_rpc_caller", None) is not None:
+            proxy._rpc_caller = self._rpc_caller
+        return proxy
+
+    def _get_rpc(self) -> Any:
+        if self._rpc_caller is None:
+            from pyisolate._internal.rpc_protocol import get_child_rpc_instance
+
+            rpc = get_child_rpc_instance()
+            if rpc is not None:
+                self._rpc_caller = rpc.create_caller(
+                    self._registry_class, self._registry_class.get_remote_id()
+                )
+            else:
+                self._rpc_caller = self._registry
+        return self._rpc_caller
+
+    def get_all_callbacks(self, call_type: str = None) -> Any:
+        return self._call_rpc("get_all_callbacks", call_type)
+
+    def get_all_wrappers(self, wrapper_type: str = None) -> Any:
+        return self._call_rpc("get_all_wrappers", wrapper_type)
+
+    def _load_list(self, *args, **kwargs) -> Any:
+        return self._call_rpc("load_list_internal", *args, **kwargs)
+
+    def prepare_hook_patches_current_keyframe(
+        self, t: Any, hook_group: Any, model_options: Any
+    ) -> None:
+        self._call_rpc(
+            "prepare_hook_patches_current_keyframe", t, hook_group, model_options
+        )
+
+    def add_hook_patches(
+        self,
+        hook: Any,
+        patches: Any,
+        strength_patch: float = 1.0,
+        strength_model: float = 1.0,
+    ) -> None:
+        self._call_rpc(
+            "add_hook_patches", hook, patches, strength_patch, strength_model
+        )
+
+    def clear_cached_hook_weights(self) -> None:
+        self._call_rpc("clear_cached_hook_weights")
+
+    def get_combined_hook_patches(self, hooks: Any) -> Any:
+        return self._call_rpc("get_combined_hook_patches", hooks)
+
+    def get_additional_models_with_key(self, key: str) -> Any:
+        return self._call_rpc("get_additional_models_with_key", key)
+
+    @property
+    def object_patches(self) -> Any:
+        return self._call_rpc("get_object_patches")
+
+    @property
+    def patches(self) -> Any:
+        res = self._call_rpc("get_patches")
+        if isinstance(res, dict):
+            new_res = {}
+            for k, v in res.items():
+                new_list = []
+                for item in v:
+                    if isinstance(item, list):
+                        new_list.append(tuple(item))
+                    else:
+                        new_list.append(item)
+                new_res[k] = new_list
+            return new_res
+        return res
+
+    @property
+    def pinned(self) -> Set:
+        val = self._call_rpc("get_patcher_attr", "pinned")
+        return set(val) if val is not None else set()
+
+    @property
+    def hook_patches(self) -> Dict:
+        val = self._call_rpc("get_patcher_attr", "hook_patches")
+        if val is None:
+            return {}
+        try:
+            from comfy.hooks import _HookRef
+            import json
+
+            new_val = {}
+            for k, v in val.items():
+                if isinstance(k, str):
+                    if k.startswith("PYISOLATE_HOOKREF:"):
+                        ref_id = k.split(":", 1)[1]
+                        h = _HookRef()
+                        h._pyisolate_id = ref_id
+                        new_val[h] = v
+                    elif k.startswith("__pyisolate_key__"):
+                        try:
+                            json_str = k[len("__pyisolate_key__") :]
+                            data = json.loads(json_str)
+                            ref_id = None
+                            if isinstance(data, list):
+                                for item in data:
+                                    if (
+                                        isinstance(item, list)
+                                        and len(item) == 2
+                                        and item[0] == "id"
+                                    ):
+                                        ref_id = item[1]
+                                        break
+                            if ref_id:
+                                h = _HookRef()
+                                h._pyisolate_id = ref_id
+                                new_val[h] = v
+                            else:
+                                new_val[k] = v
+                        except Exception:
+                            new_val[k] = v
+                    else:
+                        new_val[k] = v
+                else:
+                    new_val[k] = v
+            return new_val
+        except ImportError:
+            return val
+
+    def set_hook_mode(self, hook_mode: Any) -> None:
+        self._call_rpc("set_hook_mode", hook_mode)
+
+    def register_all_hook_patches(
+        self,
+        hooks: Any,
+        target_dict: Any,
+        model_options: Any = None,
+        registered: Any = None,
+    ) -> None:
+        self._call_rpc(
+            "register_all_hook_patches", hooks, target_dict, model_options, registered
+        )
+
+    def is_clone(self, other: Any) -> bool:
+        if isinstance(other, ModelPatcherProxy):
+            return self._call_rpc("is_clone_by_id", other._instance_id)
+        return False
+
+    def clone(self) -> ModelPatcherProxy:
+        new_id = self._call_rpc("clone")
+        return self._spawn_related_proxy(new_id)
+
+    def clone_has_same_weights(self, clone: Any) -> bool:
+        if isinstance(clone, ModelPatcherProxy):
+            return self._call_rpc("clone_has_same_weights_by_id", clone._instance_id)
+        if not IS_CHILD_PROCESS:
+            return self._call_rpc("is_clone", clone)
+        return False
+
+    def get_model_object(self, name: str) -> Any:
+        return self._call_rpc("get_model_object", name)
+
+    @property
+    def model_options(self) -> dict:
+        data = self._call_rpc("get_model_options")
+        import json
+
+        def _decode_keys(obj):
+            if isinstance(obj, dict):
+                new_d = {}
+                for k, v in obj.items():
+                    if isinstance(k, str) and k.startswith("__pyisolate_key__"):
+                        try:
+                            json_str = k[17:]
+                            val = json.loads(json_str)
+                            if isinstance(val, list):
+                                val = tuple(val)
+                            new_d[val] = _decode_keys(v)
+                        except:
+                            new_d[k] = _decode_keys(v)
+                    else:
+                        new_d[k] = _decode_keys(v)
+                return new_d
+            if isinstance(obj, list):
+                return [_decode_keys(x) for x in obj]
+            return obj
+
+        return _decode_keys(data)
+
+    @model_options.setter
+    def model_options(self, value: dict) -> None:
+        self._call_rpc("set_model_options", value)
+
+    def apply_hooks(self, hooks: Any) -> Any:
+        return self._call_rpc("apply_hooks", hooks)
+
+    def prepare_state(self, timestep: Any) -> Any:
+        return self._call_rpc("prepare_state", timestep)
+
+    def restore_hook_patches(self) -> None:
+        self._call_rpc("restore_hook_patches")
+
+    def unpatch_hooks(self, whitelist_keys_set: Optional[Set[str]] = None) -> None:
+        self._call_rpc("unpatch_hooks", whitelist_keys_set)
+
+    def model_patches_to(self, device: Any) -> Any:
+        return self._call_rpc("model_patches_to", device)
+
+    def partially_load(
+        self, device: Any, extra_memory: Any, force_patch_weights: bool = False
+    ) -> Any:
+        return self._call_rpc(
+            "partially_load", device, extra_memory, force_patch_weights
+        )
+
+    def partially_unload(
+        self, device_to: Any, memory_to_free: int = 0, force_patch_weights: bool = False
+    ) -> int:
+        return self._call_rpc(
+            "partially_unload", device_to, memory_to_free, force_patch_weights
+        )
+
+    def load(
+        self,
+        device_to: Any = None,
+        lowvram_model_memory: int = 0,
+        force_patch_weights: bool = False,
+        full_load: bool = False,
+    ) -> None:
+        self._call_rpc(
+            "load", device_to, lowvram_model_memory, force_patch_weights, full_load
+        )
+
+    def patch_model(
+        self,
+        device_to: Any = None,
+        lowvram_model_memory: int = 0,
+        load_weights: bool = True,
+        force_patch_weights: bool = False,
+    ) -> Any:
+        self._call_rpc(
+            "patch_model",
+            device_to,
+            lowvram_model_memory,
+            load_weights,
+            force_patch_weights,
+        )
+        return self
+
+    def unpatch_model(
+        self, device_to: Any = None, unpatch_weights: bool = True
+    ) -> None:
+        self._call_rpc("unpatch_model", device_to, unpatch_weights)
+
+    def detach(self, unpatch_all: bool = True) -> Any:
+        self._call_rpc("detach", unpatch_all)
+        return self.model
+
+    def _cpu_tensor_bytes(self, obj: Any) -> int:
+        import torch
+
+        if isinstance(obj, torch.Tensor):
+            if obj.device.type == "cpu":
+                return obj.nbytes
+            return 0
+        if isinstance(obj, dict):
+            return sum(self._cpu_tensor_bytes(v) for v in obj.values())
+        if isinstance(obj, (list, tuple)):
+            return sum(self._cpu_tensor_bytes(v) for v in obj)
+        return 0
+
+    def _ensure_apply_model_headroom(self, required_bytes: int) -> bool:
+        if required_bytes <= 0:
+            return True
+
+        import torch
+        import comfy.model_management as model_management
+
+        target_raw = self.load_device
+        try:
+            if isinstance(target_raw, torch.device):
+                target = target_raw
+            elif isinstance(target_raw, str):
+                target = torch.device(target_raw)
+            elif isinstance(target_raw, int):
+                target = torch.device(f"cuda:{target_raw}")
+            else:
+                target = torch.device(target_raw)
+        except Exception:
+            return True
+
+        if target.type != "cuda":
+            return True
+
+        required = required_bytes + self._APPLY_MODEL_GUARD_PADDING_BYTES
+        if model_management.get_free_memory(target) >= required:
+            return True
+
+        model_management.cleanup_models_gc()
+        model_management.cleanup_models()
+        model_management.soft_empty_cache()
+
+        if model_management.get_free_memory(target) < required:
+            model_management.free_memory(required, target, for_dynamic=True)
+            model_management.soft_empty_cache()
+
+        if model_management.get_free_memory(target) < required:
+            # Escalate to non-dynamic unloading before dispatching CUDA transfer.
+            model_management.free_memory(required, target, for_dynamic=False)
+            model_management.soft_empty_cache()
+
+        if model_management.get_free_memory(target) < required:
+            model_management.load_models_gpu(
+                [self],
+                minimum_memory_required=required,
+            )
+
+        return model_management.get_free_memory(target) >= required
+
+    def apply_model(self, *args, **kwargs) -> Any:
+        import torch
+
+        def _preferred_device() -> Any:
+            for value in args:
+                if isinstance(value, torch.Tensor):
+                    return value.device
+            for value in kwargs.values():
+                if isinstance(value, torch.Tensor):
+                    return value.device
+            return None
+
+        def _move_result_to_device(obj: Any, device: Any) -> Any:
+            if device is None:
+                return obj
+            if isinstance(obj, torch.Tensor):
+                return obj.to(device) if obj.device != device else obj
+            if isinstance(obj, dict):
+                return {k: _move_result_to_device(v, device) for k, v in obj.items()}
+            if isinstance(obj, list):
+                return [_move_result_to_device(v, device) for v in obj]
+            if isinstance(obj, tuple):
+                return tuple(_move_result_to_device(v, device) for v in obj)
+            return obj
+
+        # DynamicVRAM models must keep load/offload decisions in host process.
+        # Child-side CUDA staging here can deadlock before first inference RPC.
+        if self.is_dynamic():
+            out = self._call_rpc("inner_model_apply_model", args, kwargs)
+            return _move_result_to_device(out, _preferred_device())
+
+        required_bytes = self._cpu_tensor_bytes(args) + self._cpu_tensor_bytes(kwargs)
+        self._ensure_apply_model_headroom(required_bytes)
+        target_device = self.load_device
+
+        def _to_cuda(obj: Any) -> Any:
+            if isinstance(obj, torch.Tensor) and obj.device.type == "cpu":
+                return obj.to(target_device)
+            if isinstance(obj, dict):
+                return {k: _to_cuda(v) for k, v in obj.items()}
+            if isinstance(obj, list):
+                return [_to_cuda(v) for v in obj]
+            if isinstance(obj, tuple):
+                return tuple(_to_cuda(v) for v in obj)
+            return obj
+
+        try:
+            args_cuda = _to_cuda(args)
+            kwargs_cuda = _to_cuda(kwargs)
+        except torch.OutOfMemoryError:
+            self._ensure_apply_model_headroom(required_bytes)
+            args_cuda = _to_cuda(args)
+            kwargs_cuda = _to_cuda(kwargs)
+
+        out = self._call_rpc("inner_model_apply_model", args_cuda, kwargs_cuda)
+        return _move_result_to_device(out, _preferred_device())
+
+    def model_state_dict(self, filter_prefix: Optional[str] = None) -> Any:
+        keys = self._call_rpc("model_state_dict", filter_prefix)
+        return dict.fromkeys(keys, None)
+
+    def add_patches(self, *args: Any, **kwargs: Any) -> Any:
+        res = self._call_rpc("add_patches", *args, **kwargs)
+        if isinstance(res, list):
+            return [tuple(x) if isinstance(x, list) else x for x in res]
+        return res
+
+    def get_key_patches(self, filter_prefix: Optional[str] = None) -> Any:
+        return self._call_rpc("get_key_patches", filter_prefix)
+
+    def patch_weight_to_device(self, key, device_to=None, inplace_update=False):
+        self._call_rpc("patch_weight_to_device", key, device_to, inplace_update)
+
+    def pin_weight_to_device(self, key):
+        self._call_rpc("pin_weight_to_device", key)
+
+    def unpin_weight(self, key):
+        self._call_rpc("unpin_weight", key)
+
+    def unpin_all_weights(self):
+        self._call_rpc("unpin_all_weights")
+
+    def calculate_weight(self, patches, weight, key, intermediate_dtype=None):
+        return self._call_rpc(
+            "calculate_weight", patches, weight, key, intermediate_dtype
+        )
+
+    def inject_model(self) -> None:
+        self._call_rpc("inject_model")
+
+    def eject_model(self) -> None:
+        self._call_rpc("eject_model")
+
+    def use_ejected(self, skip_and_inject_on_exit_only: bool = False) -> Any:
+        return AutoPatcherEjector(
+            self, skip_and_inject_on_exit_only=skip_and_inject_on_exit_only
+        )
+
+    @property
+    def is_injected(self) -> bool:
+        return self._call_rpc("get_is_injected")
+
+    @property
+    def skip_injection(self) -> bool:
+        return self._call_rpc("get_skip_injection")
+
+    @skip_injection.setter
+    def skip_injection(self, value: bool) -> None:
+        self._call_rpc("set_skip_injection", value)
+
+    def clean_hooks(self) -> None:
+        self._call_rpc("clean_hooks")
+
+    def pre_run(self) -> None:
+        self._call_rpc("pre_run")
+
+    def cleanup(self) -> None:
+        try:
+            self._call_rpc("cleanup")
+        except Exception:
+            logger.debug(
+                "ModelPatcherProxy cleanup RPC failed for %s",
+                self._instance_id,
+                exc_info=True,
+            )
+        finally:
+            super().cleanup()
+
+    @property
+    def model(self) -> _InnerModelProxy:
+        return _InnerModelProxy(self)
+
+    def __getattr__(self, name: str) -> Any:
+        _whitelisted_attrs = {
+            "hook_patches_backup",
+            "hook_backup",
+            "cached_hook_patches",
+            "current_hooks",
+            "forced_hooks",
+            "is_clip",
+            "patches_uuid",
+            "pinned",
+            "attachments",
+            "additional_models",
+            "injections",
+            "hook_patches",
+            "model_lowvram",
+            "model_loaded_weight_memory",
+            "backup",
+            "object_patches_backup",
+            "weight_wrapper_patches",
+            "weight_inplace_update",
+            "force_cast_weights",
+        }
+        if name in _whitelisted_attrs:
+            return self._call_rpc("get_patcher_attr", name)
+        raise AttributeError(
+            f"'{type(self).__name__}' object has no attribute '{name}'"
+        )
+
+    def load_lora(
+        self,
+        lora_path: str,
+        strength_model: float,
+        clip: Optional[Any] = None,
+        strength_clip: float = 1.0,
+    ) -> tuple:
+        clip_id = None
+        if clip is not None:
+            clip_id = getattr(clip, "_instance_id", getattr(clip, "_clip_id", None))
+        result = self._call_rpc(
+            "load_lora", lora_path, strength_model, clip_id, strength_clip
+        )
+        new_model = None
+        if result.get("model_id"):
+            new_model = self._spawn_related_proxy(result["model_id"])
+        new_clip = None
+        if result.get("clip_id"):
+            from comfy.isolation.clip_proxy import CLIPProxy
+
+            new_clip = CLIPProxy(result["clip_id"])
+        return (new_model, new_clip)
+
+    @property
+    def load_device(self) -> Any:
+        return self._call_rpc("get_load_device")
+
+    @property
+    def offload_device(self) -> Any:
+        return self._call_rpc("get_offload_device")
+
+    @property
+    def device(self) -> Any:
+        return self.load_device
+
+    def current_loaded_device(self) -> Any:
+        return self._call_rpc("current_loaded_device")
+
+    @property
+    def size(self) -> int:
+        return self._call_rpc("get_size")
+
+    def model_size(self) -> Any:
+        return self._call_rpc("model_size")
+
+    def loaded_size(self) -> Any:
+        return self._call_rpc("loaded_size")
+
+    def get_ram_usage(self) -> int:
+        return self._call_rpc("get_ram_usage")
+
+    def lowvram_patch_counter(self) -> int:
+        return self._call_rpc("lowvram_patch_counter")
+
+    def memory_required(self, input_shape: Any) -> Any:
+        return self._call_rpc("memory_required", input_shape)
+
+    def get_operation_state(self) -> Dict[str, Any]:
+        state = self._call_rpc("get_operation_state")
+        return state if isinstance(state, dict) else {}
+
+    def wait_for_idle(self, timeout_ms: int = 0) -> bool:
+        return bool(self._call_rpc("wait_for_idle", timeout_ms))
+
+    def is_dynamic(self) -> bool:
+        return bool(self._call_rpc("is_dynamic"))
+
+    def get_free_memory(self, device: Any) -> Any:
+        return self._call_rpc("get_free_memory", device)
+
+    def partially_unload_ram(self, ram_to_unload: int) -> Any:
+        return self._call_rpc("partially_unload_ram", ram_to_unload)
+
+    def model_dtype(self) -> Any:
+        res = self._call_rpc("model_dtype")
+        if isinstance(res, str) and res.startswith("torch."):
+            try:
+                import torch
+
+                attr = res.split(".")[-1]
+                if hasattr(torch, attr):
+                    return getattr(torch, attr)
+            except ImportError:
+                pass
+        return res
+
+    @property
+    def hook_mode(self) -> Any:
+        return self._call_rpc("get_hook_mode")
+
+    @hook_mode.setter
+    def hook_mode(self, value: Any) -> None:
+        self._call_rpc("set_hook_mode", value)
+
+    def set_model_sampler_cfg_function(
+        self, sampler_cfg_function: Any, disable_cfg1_optimization: bool = False
+    ) -> None:
+        self._call_rpc(
+            "set_model_sampler_cfg_function",
+            sampler_cfg_function,
+            disable_cfg1_optimization,
+        )
+
+    def set_model_sampler_post_cfg_function(
+        self, post_cfg_function: Any, disable_cfg1_optimization: bool = False
+    ) -> None:
+        self._call_rpc(
+            "set_model_sampler_post_cfg_function",
+            post_cfg_function,
+            disable_cfg1_optimization,
+        )
+
+    def set_model_sampler_pre_cfg_function(
+        self, pre_cfg_function: Any, disable_cfg1_optimization: bool = False
+    ) -> None:
+        self._call_rpc(
+            "set_model_sampler_pre_cfg_function",
+            pre_cfg_function,
+            disable_cfg1_optimization,
+        )
+
+    def set_model_sampler_calc_cond_batch_function(self, fn: Any) -> None:
+        self._call_rpc("set_model_sampler_calc_cond_batch_function", fn)
+
+    def set_model_unet_function_wrapper(self, unet_wrapper_function: Any) -> None:
+        self._call_rpc("set_model_unet_function_wrapper", unet_wrapper_function)
+
+    def set_model_denoise_mask_function(self, denoise_mask_function: Any) -> None:
+        self._call_rpc("set_model_denoise_mask_function", denoise_mask_function)
+
+    def set_model_patch(self, patch: Any, name: str) -> None:
+        self._call_rpc("set_model_patch", patch, name)
+
+    def set_model_patch_replace(
+        self,
+        patch: Any,
+        name: str,
+        block_name: str,
+        number: int,
+        transformer_index: Optional[int] = None,
+    ) -> None:
+        self._call_rpc(
+            "set_model_patch_replace",
+            patch,
+            name,
+            block_name,
+            number,
+            transformer_index,
+        )
+
+    def set_model_attn1_patch(self, patch: Any) -> None:
+        self.set_model_patch(patch, "attn1_patch")
+
+    def set_model_attn2_patch(self, patch: Any) -> None:
+        self.set_model_patch(patch, "attn2_patch")
+
+    def set_model_attn1_replace(
+        self,
+        patch: Any,
+        block_name: str,
+        number: int,
+        transformer_index: Optional[int] = None,
+    ) -> None:
+        self.set_model_patch_replace(
+            patch, "attn1", block_name, number, transformer_index
+        )
+
+    def set_model_attn2_replace(
+        self,
+        patch: Any,
+        block_name: str,
+        number: int,
+        transformer_index: Optional[int] = None,
+    ) -> None:
+        self.set_model_patch_replace(
+            patch, "attn2", block_name, number, transformer_index
+        )
+
+    def set_model_attn1_output_patch(self, patch: Any) -> None:
+        self.set_model_patch(patch, "attn1_output_patch")
+
+    def set_model_attn2_output_patch(self, patch: Any) -> None:
+        self.set_model_patch(patch, "attn2_output_patch")
+
+    def set_model_input_block_patch(self, patch: Any) -> None:
+        self.set_model_patch(patch, "input_block_patch")
+
+    def set_model_input_block_patch_after_skip(self, patch: Any) -> None:
+        self.set_model_patch(patch, "input_block_patch_after_skip")
+
+    def set_model_output_block_patch(self, patch: Any) -> None:
+        self.set_model_patch(patch, "output_block_patch")
+
+    def set_model_emb_patch(self, patch: Any) -> None:
+        self.set_model_patch(patch, "emb_patch")
+
+    def set_model_forward_timestep_embed_patch(self, patch: Any) -> None:
+        self.set_model_patch(patch, "forward_timestep_embed_patch")
+
+    def set_model_double_block_patch(self, patch: Any) -> None:
+        self.set_model_patch(patch, "double_block")
+
+    def set_model_post_input_patch(self, patch: Any) -> None:
+        self.set_model_patch(patch, "post_input")
+
+    def set_model_rope_options(
+        self,
+        scale_x=1.0,
+        shift_x=0.0,
+        scale_y=1.0,
+        shift_y=0.0,
+        scale_t=1.0,
+        shift_t=0.0,
+        **kwargs: Any,
+    ) -> None:
+        options = {
+            "scale_x": scale_x,
+            "shift_x": shift_x,
+            "scale_y": scale_y,
+            "shift_y": shift_y,
+            "scale_t": scale_t,
+            "shift_t": shift_t,
+        }
+        options.update(kwargs)
+        self._call_rpc("set_model_rope_options", options)
+
+    def set_model_compute_dtype(self, dtype: Any) -> None:
+        self._call_rpc("set_model_compute_dtype", dtype)
+
+    def add_object_patch(self, name: str, obj: Any) -> None:
+        self._call_rpc("add_object_patch", name, obj)
+
+    def add_weight_wrapper(self, name: str, function: Any) -> None:
+        self._call_rpc("add_weight_wrapper", name, function)
+
+    def add_wrapper_with_key(self, wrapper_type: Any, key: str, fn: Any) -> None:
+        self._call_rpc("add_wrapper_with_key", wrapper_type, key, fn)
+
+    def add_wrapper(self, wrapper_type: str, wrapper: Callable) -> None:
+        self.add_wrapper_with_key(wrapper_type, None, wrapper)
+
+    def remove_wrappers_with_key(self, wrapper_type: str, key: str) -> None:
+        self._call_rpc("remove_wrappers_with_key", wrapper_type, key)
+
+    @property
+    def wrappers(self) -> Any:
+        return self._call_rpc("get_wrappers")
+
+    def add_callback_with_key(self, call_type: str, key: str, callback: Any) -> None:
+        self._call_rpc("add_callback_with_key", call_type, key, callback)
+
+    def add_callback(self, call_type: str, callback: Any) -> None:
+        self.add_callback_with_key(call_type, None, callback)
+
+    def remove_callbacks_with_key(self, call_type: str, key: str) -> None:
+        self._call_rpc("remove_callbacks_with_key", call_type, key)
+
+    @property
+    def callbacks(self) -> Any:
+        return self._call_rpc("get_callbacks")
+
+    def set_attachments(self, key: str, attachment: Any) -> None:
+        self._call_rpc("set_attachments", key, attachment)
+
+    def get_attachment(self, key: str) -> Any:
+        return self._call_rpc("get_attachment", key)
+
+    def remove_attachments(self, key: str) -> None:
+        self._call_rpc("remove_attachments", key)
+
+    def set_injections(self, key: str, injections: Any) -> None:
+        self._call_rpc("set_injections", key, injections)
+
+    def get_injections(self, key: str) -> Any:
+        return self._call_rpc("get_injections", key)
+
+    def remove_injections(self, key: str) -> None:
+        self._call_rpc("remove_injections", key)
+
+    def set_additional_models(self, key: str, models: Any) -> None:
+        ids = [m._instance_id for m in models]
+        self._call_rpc("set_additional_models", key, ids)
+
+    def remove_additional_models(self, key: str) -> None:
+        self._call_rpc("remove_additional_models", key)
+
+    def get_nested_additional_models(self) -> Any:
+        return self._call_rpc("get_nested_additional_models")
+
+    def get_additional_models(self) -> List[ModelPatcherProxy]:
+        ids = self._call_rpc("get_additional_models")
+        return [self._spawn_related_proxy(mid) for mid in ids]
+
+    def model_patches_models(self) -> Any:
+        return self._call_rpc("model_patches_models")
+
+    @property
+    def parent(self) -> Any:
+        return self._call_rpc("get_parent")
+
+    def model_mmap_residency(self, free: bool = False) -> tuple:
+        result = self._call_rpc("model_mmap_residency", free)
+        if isinstance(result, list):
+            return tuple(result)
+        return result
+
+    def pinned_memory_size(self) -> int:
+        return self._call_rpc("pinned_memory_size")
+
+    def get_non_dynamic_delegate(self) -> ModelPatcherProxy:
+        new_id = self._call_rpc("get_non_dynamic_delegate")
+        return self._spawn_related_proxy(new_id)
+
+    def disable_model_cfg1_optimization(self) -> None:
+        self._call_rpc("disable_model_cfg1_optimization")
+
+    def set_model_noise_refiner_patch(self, patch: Any) -> None:
+        self.set_model_patch(patch, "noise_refiner")
+
+
+class _InnerModelProxy:
+    def __init__(self, parent: ModelPatcherProxy):
+        self._parent = parent
+        self._model_sampling = None
+
+    def __getattr__(self, name: str) -> Any:
+        if name.startswith("_"):
+            raise AttributeError(name)
+        if name == "model_config":
+            from types import SimpleNamespace
+
+            data = self._parent._call_rpc("get_inner_model_attr", name)
+            if isinstance(data, dict):
+                return SimpleNamespace(**data)
+            return data
+        if name in (
+            "latent_format",
+            "model_type",
+            "current_weight_patches_uuid",
+        ):
+            return self._parent._call_rpc("get_inner_model_attr", name)
+        if name == "load_device":
+            return self._parent._call_rpc("get_inner_model_attr", "load_device")
+        if name == "device":
+            return self._parent._call_rpc("get_inner_model_attr", "device")
+        if name == "current_patcher":
+            proxy = ModelPatcherProxy(
+                self._parent._instance_id,
+                self._parent._registry,
+                manage_lifecycle=False,
+            )
+            if getattr(self._parent, "_rpc_caller", None) is not None:
+                proxy._rpc_caller = self._parent._rpc_caller
+            return proxy
+        if name == "model_sampling":
+            if self._model_sampling is None:
+                self._model_sampling = self._parent._call_rpc(
+                    "get_model_object", "model_sampling"
+                )
+            return self._model_sampling
+        if name == "extra_conds_shapes":
+            return lambda *a, **k: self._parent._call_rpc(
+                "inner_model_extra_conds_shapes", a, k
+            )
+        if name == "extra_conds":
+            return lambda *a, **k: self._parent._call_rpc(
+                "inner_model_extra_conds", a, k
+            )
+        if name == "memory_required":
+            return lambda *a, **k: self._parent._call_rpc(
+                "inner_model_memory_required", a, k
+            )
+        if name == "apply_model":
+            # Delegate to parent's method to get the CPU->CUDA optimization
+            return self._parent.apply_model
+        if name == "process_latent_in":
+            return lambda *a, **k: self._parent._call_rpc("process_latent_in", a, k)
+        if name == "process_latent_out":
+            return lambda *a, **k: self._parent._call_rpc("process_latent_out", a, k)
+        if name == "scale_latent_inpaint":
+            return lambda *a, **k: self._parent._call_rpc("scale_latent_inpaint", a, k)
+        if name == "diffusion_model":
+            return self._parent._call_rpc("get_inner_model_attr", "diffusion_model")
+        if name == "state_dict":
+            return lambda: self._parent.model_state_dict()
+        raise AttributeError(f"'{name}' not supported on isolated InnerModel")

comfy/isolation/model_patcher_proxy_utils.py

@@ -0,0 +1,150 @@
+# pylint: disable=import-outside-toplevel,logging-fstring-interpolation,protected-access
+# Isolation utilities and serializers for ModelPatcherProxy
+from __future__ import annotations
+
+import logging
+import os
+from typing import Any
+
+from comfy.cli_args import args
+
+logger = logging.getLogger(__name__)
+
+
+def maybe_wrap_model_for_isolation(model_patcher: Any) -> Any:
+    from comfy.isolation.model_patcher_proxy_registry import ModelPatcherRegistry
+    from comfy.isolation.model_patcher_proxy import ModelPatcherProxy
+
+    is_child = os.environ.get("PYISOLATE_CHILD") == "1"
+    isolation_active = args.use_process_isolation or is_child
+
+    if not isolation_active:
+        return model_patcher
+    if is_child:
+        return model_patcher
+    if isinstance(model_patcher, ModelPatcherProxy):
+        return model_patcher
+
+    registry = ModelPatcherRegistry()
+    model_id = registry.register(model_patcher)
+    logger.debug(f"Isolated ModelPatcher: {model_id}")
+    return ModelPatcherProxy(model_id, registry, manage_lifecycle=True)
+
+
+def register_hooks_serializers(registry=None):
+    from pyisolate._internal.serialization_registry import SerializerRegistry
+    import comfy.hooks
+
+    if registry is None:
+        registry = SerializerRegistry.get_instance()
+
+    def serialize_enum(obj):
+        return {"__enum__": f"{type(obj).__name__}.{obj.name}"}
+
+    def deserialize_enum(data):
+        cls_name, val_name = data["__enum__"].split(".")
+        cls = getattr(comfy.hooks, cls_name)
+        return cls[val_name]
+
+    registry.register("EnumHookType", serialize_enum, deserialize_enum)
+    registry.register("EnumHookScope", serialize_enum, deserialize_enum)
+    registry.register("EnumHookMode", serialize_enum, deserialize_enum)
+    registry.register("EnumWeightTarget", serialize_enum, deserialize_enum)
+
+    def serialize_hook_group(obj):
+        return {"__type__": "HookGroup", "hooks": obj.hooks}
+
+    def deserialize_hook_group(data):
+        hg = comfy.hooks.HookGroup()
+        for h in data["hooks"]:
+            hg.add(h)
+        return hg
+
+    registry.register("HookGroup", serialize_hook_group, deserialize_hook_group)
+
+    def serialize_dict_state(obj):
+        d = obj.__dict__.copy()
+        d["__type__"] = type(obj).__name__
+        if "custom_should_register" in d:
+            del d["custom_should_register"]
+        return d
+
+    def deserialize_dict_state_generic(cls):
+        def _deserialize(data):
+            h = cls()
+            h.__dict__.update(data)
+            return h
+
+        return _deserialize
+
+    def deserialize_hook_keyframe(data):
+        h = comfy.hooks.HookKeyframe(strength=data.get("strength", 1.0))
+        h.__dict__.update(data)
+        return h
+
+    registry.register("HookKeyframe", serialize_dict_state, deserialize_hook_keyframe)
+
+    def deserialize_hook_keyframe_group(data):
+        h = comfy.hooks.HookKeyframeGroup()
+        h.__dict__.update(data)
+        return h
+
+    registry.register(
+        "HookKeyframeGroup", serialize_dict_state, deserialize_hook_keyframe_group
+    )
+
+    def deserialize_hook(data):
+        h = comfy.hooks.Hook()
+        h.__dict__.update(data)
+        return h
+
+    registry.register("Hook", serialize_dict_state, deserialize_hook)
+
+    def deserialize_weight_hook(data):
+        h = comfy.hooks.WeightHook()
+        h.__dict__.update(data)
+        return h
+
+    registry.register("WeightHook", serialize_dict_state, deserialize_weight_hook)
+
+    def serialize_set(obj):
+        return {"__set__": list(obj)}
+
+    def deserialize_set(data):
+        return set(data["__set__"])
+
+    registry.register("set", serialize_set, deserialize_set)
+
+    try:
+        from comfy.weight_adapter.lora import LoRAAdapter
+
+        def serialize_lora(obj):
+            return {"weights": {}, "loaded_keys": list(obj.loaded_keys)}
+
+        def deserialize_lora(data):
+            return LoRAAdapter(set(data["loaded_keys"]), data["weights"])
+
+        registry.register("LoRAAdapter", serialize_lora, deserialize_lora)
+    except Exception:
+        pass
+
+    try:
+        from comfy.hooks import _HookRef
+        import uuid
+
+        def serialize_hook_ref(obj):
+            return {
+                "__hook_ref__": True,
+                "id": getattr(obj, "_pyisolate_id", str(uuid.uuid4())),
+            }
+
+        def deserialize_hook_ref(data):
+            h = _HookRef()
+            h._pyisolate_id = data.get("id", str(uuid.uuid4()))
+            return h
+
+        registry.register("_HookRef", serialize_hook_ref, deserialize_hook_ref)
+    except ImportError:
+        pass
+    except Exception as e:
+        logger.warning(f"Failed to register _HookRef: {e}")

comfy/isolation/model_sampling_proxy.py

@@ -0,0 +1,360 @@
+# pylint: disable=import-outside-toplevel
+from __future__ import annotations
+
+import asyncio
+import logging
+import os
+import threading
+import time
+from typing import Any
+
+from comfy.isolation.proxies.base import (
+    BaseProxy,
+    BaseRegistry,
+    detach_if_grad,
+    get_thread_loop,
+    run_coro_in_new_loop,
+)
+
+logger = logging.getLogger(__name__)
+
+
+def _describe_value(obj: Any) -> str:
+    try:
+        import torch
+    except Exception:
+        torch = None
+    try:
+        if torch is not None and isinstance(obj, torch.Tensor):
+            return (
+                "Tensor(shape=%s,dtype=%s,device=%s,id=%s)"
+                % (tuple(obj.shape), obj.dtype, obj.device, id(obj))
+            )
+    except Exception:
+        pass
+    return "%s(id=%s)" % (type(obj).__name__, id(obj))
+
+
+def _prefer_device(*tensors: Any) -> Any:
+    try:
+        import torch
+    except Exception:
+        return None
+    for t in tensors:
+        if isinstance(t, torch.Tensor) and t.is_cuda:
+            return t.device
+    for t in tensors:
+        if isinstance(t, torch.Tensor):
+            return t.device
+    return None
+
+
+def _to_device(obj: Any, device: Any) -> Any:
+    try:
+        import torch
+    except Exception:
+        return obj
+    if device is None:
+        return obj
+    if isinstance(obj, torch.Tensor):
+        if obj.device != device:
+            return obj.to(device)
+        return obj
+    if isinstance(obj, (list, tuple)):
+        converted = [_to_device(x, device) for x in obj]
+        return type(obj)(converted) if isinstance(obj, tuple) else converted
+    if isinstance(obj, dict):
+        return {k: _to_device(v, device) for k, v in obj.items()}
+    return obj
+
+
+def _to_cpu_for_rpc(obj: Any) -> Any:
+    try:
+        import torch
+    except Exception:
+        return obj
+    if isinstance(obj, torch.Tensor):
+        t = obj.detach() if obj.requires_grad else obj
+        if t.is_cuda:
+            return t.to("cpu")
+        return t
+    if isinstance(obj, (list, tuple)):
+        converted = [_to_cpu_for_rpc(x) for x in obj]
+        return type(obj)(converted) if isinstance(obj, tuple) else converted
+    if isinstance(obj, dict):
+        return {k: _to_cpu_for_rpc(v) for k, v in obj.items()}
+    return obj
+
+
+class ModelSamplingRegistry(BaseRegistry[Any]):
+    _type_prefix = "modelsampling"
+
+    async def calculate_input(self, instance_id: str, sigma: Any, noise: Any) -> Any:
+        sampling = self._get_instance(instance_id)
+        return detach_if_grad(sampling.calculate_input(sigma, noise))
+
+    async def calculate_denoised(
+        self, instance_id: str, sigma: Any, model_output: Any, model_input: Any
+    ) -> Any:
+        sampling = self._get_instance(instance_id)
+        return detach_if_grad(
+            sampling.calculate_denoised(sigma, model_output, model_input)
+        )
+
+    async def noise_scaling(
+        self,
+        instance_id: str,
+        sigma: Any,
+        noise: Any,
+        latent_image: Any,
+        max_denoise: bool = False,
+    ) -> Any:
+        sampling = self._get_instance(instance_id)
+        return detach_if_grad(
+            sampling.noise_scaling(sigma, noise, latent_image, max_denoise=max_denoise)
+        )
+
+    async def inverse_noise_scaling(
+        self, instance_id: str, sigma: Any, latent: Any
+    ) -> Any:
+        sampling = self._get_instance(instance_id)
+        return detach_if_grad(sampling.inverse_noise_scaling(sigma, latent))
+
+    async def timestep(self, instance_id: str, sigma: Any) -> Any:
+        sampling = self._get_instance(instance_id)
+        return sampling.timestep(sigma)
+
+    async def sigma(self, instance_id: str, timestep: Any) -> Any:
+        sampling = self._get_instance(instance_id)
+        return sampling.sigma(timestep)
+
+    async def percent_to_sigma(self, instance_id: str, percent: float) -> Any:
+        sampling = self._get_instance(instance_id)
+        return sampling.percent_to_sigma(percent)
+
+    async def get_sigma_min(self, instance_id: str) -> Any:
+        sampling = self._get_instance(instance_id)
+        return detach_if_grad(sampling.sigma_min)
+
+    async def get_sigma_max(self, instance_id: str) -> Any:
+        sampling = self._get_instance(instance_id)
+        return detach_if_grad(sampling.sigma_max)
+
+    async def get_sigma_data(self, instance_id: str) -> Any:
+        sampling = self._get_instance(instance_id)
+        return detach_if_grad(sampling.sigma_data)
+
+    async def get_sigmas(self, instance_id: str) -> Any:
+        sampling = self._get_instance(instance_id)
+        return detach_if_grad(sampling.sigmas)
+
+    async def set_sigmas(self, instance_id: str, sigmas: Any) -> None:
+        sampling = self._get_instance(instance_id)
+        sampling.set_sigmas(sigmas)
+
+
+class ModelSamplingProxy(BaseProxy[ModelSamplingRegistry]):
+    _registry_class = ModelSamplingRegistry
+    __module__ = "comfy.isolation.model_sampling_proxy"
+
+    def _get_rpc(self) -> Any:
+        if self._rpc_caller is None:
+            from pyisolate._internal.rpc_protocol import get_child_rpc_instance
+
+            rpc = get_child_rpc_instance()
+            if rpc is not None:
+                self._rpc_caller = rpc.create_caller(
+                    ModelSamplingRegistry, ModelSamplingRegistry.get_remote_id()
+                )
+            else:
+                registry = ModelSamplingRegistry()
+
+                class _LocalCaller:
+                    def calculate_input(
+                        self, instance_id: str, sigma: Any, noise: Any
+                    ) -> Any:
+                        return registry.calculate_input(instance_id, sigma, noise)
+
+                    def calculate_denoised(
+                        self,
+                        instance_id: str,
+                        sigma: Any,
+                        model_output: Any,
+                        model_input: Any,
+                    ) -> Any:
+                        return registry.calculate_denoised(
+                            instance_id, sigma, model_output, model_input
+                        )
+
+                    def noise_scaling(
+                        self,
+                        instance_id: str,
+                        sigma: Any,
+                        noise: Any,
+                        latent_image: Any,
+                        max_denoise: bool = False,
+                    ) -> Any:
+                        return registry.noise_scaling(
+                            instance_id, sigma, noise, latent_image, max_denoise
+                        )
+
+                    def inverse_noise_scaling(
+                        self, instance_id: str, sigma: Any, latent: Any
+                    ) -> Any:
+                        return registry.inverse_noise_scaling(
+                            instance_id, sigma, latent
+                        )
+
+                    def timestep(self, instance_id: str, sigma: Any) -> Any:
+                        return registry.timestep(instance_id, sigma)
+
+                    def sigma(self, instance_id: str, timestep: Any) -> Any:
+                        return registry.sigma(instance_id, timestep)
+
+                    def percent_to_sigma(self, instance_id: str, percent: float) -> Any:
+                        return registry.percent_to_sigma(instance_id, percent)
+
+                    def get_sigma_min(self, instance_id: str) -> Any:
+                        return registry.get_sigma_min(instance_id)
+
+                    def get_sigma_max(self, instance_id: str) -> Any:
+                        return registry.get_sigma_max(instance_id)
+
+                    def get_sigma_data(self, instance_id: str) -> Any:
+                        return registry.get_sigma_data(instance_id)
+
+                    def get_sigmas(self, instance_id: str) -> Any:
+                        return registry.get_sigmas(instance_id)
+
+                    def set_sigmas(self, instance_id: str, sigmas: Any) -> None:
+                        return registry.set_sigmas(instance_id, sigmas)
+
+                self._rpc_caller = _LocalCaller()
+        return self._rpc_caller
+
+    def _call(self, method_name: str, *args: Any) -> Any:
+        rpc = self._get_rpc()
+        method = getattr(rpc, method_name)
+        result = method(self._instance_id, *args)
+        timeout_ms = self._rpc_timeout_ms()
+        start_epoch = time.time()
+        start_perf = time.perf_counter()
+        thread_id = threading.get_ident()
+        call_id = "%s:%s:%s:%.6f" % (
+            self._instance_id,
+            method_name,
+            thread_id,
+            start_perf,
+        )
+        logger.debug(
+            "ISO:modelsampling_rpc_start method=%s instance_id=%s call_id=%s start_ts=%.6f thread=%s timeout_ms=%s",
+            method_name,
+            self._instance_id,
+            call_id,
+            start_epoch,
+            thread_id,
+            timeout_ms,
+        )
+        if asyncio.iscoroutine(result):
+            result = asyncio.wait_for(result, timeout=timeout_ms / 1000.0)
+            try:
+                asyncio.get_running_loop()
+                out = run_coro_in_new_loop(result)
+            except RuntimeError:
+                loop = get_thread_loop()
+                out = loop.run_until_complete(result)
+        else:
+            out = result
+        logger.debug(
+            "ISO:modelsampling_rpc_after_await method=%s instance_id=%s call_id=%s out=%s",
+            method_name,
+            self._instance_id,
+            call_id,
+            _describe_value(out),
+        )
+        elapsed_ms = (time.perf_counter() - start_perf) * 1000.0
+        logger.debug(
+            "ISO:modelsampling_rpc_end method=%s instance_id=%s call_id=%s elapsed_ms=%.3f thread=%s",
+            method_name,
+            self._instance_id,
+            call_id,
+            elapsed_ms,
+            thread_id,
+        )
+        logger.debug(
+            "ISO:modelsampling_rpc_return method=%s instance_id=%s call_id=%s",
+            method_name,
+            self._instance_id,
+            call_id,
+        )
+        return out
+
+    @staticmethod
+    def _rpc_timeout_ms() -> int:
+        raw = os.environ.get(
+            "COMFY_ISOLATION_MODEL_SAMPLING_RPC_TIMEOUT_MS",
+            os.environ.get("COMFY_ISOLATION_LOAD_RPC_TIMEOUT_MS", "30000"),
+        )
+        try:
+            timeout_ms = int(raw)
+        except ValueError:
+            timeout_ms = 30000
+        return max(1, timeout_ms)
+
+    @property
+    def sigma_min(self) -> Any:
+        return self._call("get_sigma_min")
+
+    @property
+    def sigma_max(self) -> Any:
+        return self._call("get_sigma_max")
+
+    @property
+    def sigma_data(self) -> Any:
+        return self._call("get_sigma_data")
+
+    @property
+    def sigmas(self) -> Any:
+        return self._call("get_sigmas")
+
+    def calculate_input(self, sigma: Any, noise: Any) -> Any:
+        return self._call("calculate_input", sigma, noise)
+
+    def calculate_denoised(
+        self, sigma: Any, model_output: Any, model_input: Any
+    ) -> Any:
+        return self._call("calculate_denoised", sigma, model_output, model_input)
+
+    def noise_scaling(
+        self, sigma: Any, noise: Any, latent_image: Any, max_denoise: bool = False
+    ) -> Any:
+        preferred_device = _prefer_device(noise, latent_image)
+        out = self._call(
+            "noise_scaling",
+            _to_cpu_for_rpc(sigma),
+            _to_cpu_for_rpc(noise),
+            _to_cpu_for_rpc(latent_image),
+            max_denoise,
+        )
+        return _to_device(out, preferred_device)
+
+    def inverse_noise_scaling(self, sigma: Any, latent: Any) -> Any:
+        preferred_device = _prefer_device(latent)
+        out = self._call(
+            "inverse_noise_scaling",
+            _to_cpu_for_rpc(sigma),
+            _to_cpu_for_rpc(latent),
+        )
+        return _to_device(out, preferred_device)
+
+    def timestep(self, sigma: Any) -> Any:
+        return self._call("timestep", sigma)
+
+    def sigma(self, timestep: Any) -> Any:
+        return self._call("sigma", timestep)
+
+    def percent_to_sigma(self, percent: float) -> Any:
+        return self._call("percent_to_sigma", percent)
+
+    def set_sigmas(self, sigmas: Any) -> None:
+        return self._call("set_sigmas", sigmas)

comfy/isolation/proxies/__init__.py

@@ -0,0 +1,17 @@
+from .base import (
+    IS_CHILD_PROCESS,
+    BaseProxy,
+    BaseRegistry,
+    detach_if_grad,
+    get_thread_loop,
+    run_coro_in_new_loop,
+)
+
+__all__ = [
+    "IS_CHILD_PROCESS",
+    "BaseRegistry",
+    "BaseProxy",
+    "get_thread_loop",
+    "run_coro_in_new_loop",
+    "detach_if_grad",
+]

comfy/isolation/proxies/base.py

@@ -0,0 +1,301 @@
+# pylint: disable=global-statement,import-outside-toplevel,protected-access
+from __future__ import annotations
+
+import asyncio
+import concurrent.futures
+import logging
+import os
+import threading
+import time
+import weakref
+from typing import Any, Callable, Dict, Generic, Optional, TypeVar
+
+try:
+    from pyisolate import ProxiedSingleton
+except ImportError:
+
+    class ProxiedSingleton:  # type: ignore[no-redef]
+        pass
+
+
+logger = logging.getLogger(__name__)
+
+IS_CHILD_PROCESS = os.environ.get("PYISOLATE_CHILD") == "1"
+_thread_local = threading.local()
+T = TypeVar("T")
+
+
+def get_thread_loop() -> asyncio.AbstractEventLoop:
+    loop = getattr(_thread_local, "loop", None)
+    if loop is None or loop.is_closed():
+        loop = asyncio.new_event_loop()
+        _thread_local.loop = loop
+    return loop
+
+
+def run_coro_in_new_loop(coro: Any) -> Any:
+    result_box: Dict[str, Any] = {}
+    exc_box: Dict[str, BaseException] = {}
+
+    def runner() -> None:
+        loop = asyncio.new_event_loop()
+        asyncio.set_event_loop(loop)
+        try:
+            result_box["value"] = loop.run_until_complete(coro)
+        except Exception as exc:  # noqa: BLE001
+            exc_box["exc"] = exc
+        finally:
+            loop.close()
+
+    t = threading.Thread(target=runner, daemon=True)
+    t.start()
+    t.join()
+    if "exc" in exc_box:
+        raise exc_box["exc"]
+    return result_box.get("value")
+
+
+def detach_if_grad(obj: Any) -> Any:
+    try:
+        import torch
+    except Exception:
+        return obj
+
+    if isinstance(obj, torch.Tensor):
+        return obj.detach() if obj.requires_grad else obj
+    if isinstance(obj, (list, tuple)):
+        return type(obj)(detach_if_grad(x) for x in obj)
+    if isinstance(obj, dict):
+        return {k: detach_if_grad(v) for k, v in obj.items()}
+    return obj
+
+
+class BaseRegistry(ProxiedSingleton, Generic[T]):
+    _type_prefix: str = "base"
+
+    def __init__(self) -> None:
+        if hasattr(ProxiedSingleton, "__init__") and ProxiedSingleton is not object:
+            super().__init__()
+        self._registry: Dict[str, T] = {}
+        self._id_map: Dict[int, str] = {}
+        self._counter = 0
+        self._lock = threading.Lock()
+
+    def register(self, instance: T) -> str:
+        with self._lock:
+            obj_id = id(instance)
+            if obj_id in self._id_map:
+                return self._id_map[obj_id]
+            instance_id = f"{self._type_prefix}_{self._counter}"
+            self._counter += 1
+            self._registry[instance_id] = instance
+            self._id_map[obj_id] = instance_id
+        return instance_id
+
+    def unregister_sync(self, instance_id: str) -> None:
+        with self._lock:
+            instance = self._registry.pop(instance_id, None)
+            if instance:
+                self._id_map.pop(id(instance), None)
+
+    def _get_instance(self, instance_id: str) -> T:
+        if IS_CHILD_PROCESS:
+            raise RuntimeError(
+                f"[{self.__class__.__name__}] _get_instance called in child"
+            )
+        with self._lock:
+            instance = self._registry.get(instance_id)
+        if instance is None:
+            raise ValueError(f"{instance_id} not found")
+        return instance
+
+
+_GLOBAL_LOOP: Optional[asyncio.AbstractEventLoop] = None
+
+
+def set_global_loop(loop: asyncio.AbstractEventLoop) -> None:
+    global _GLOBAL_LOOP
+    _GLOBAL_LOOP = loop
+
+
+def run_sync_rpc_coro(coro: Any, timeout_ms: Optional[int] = None) -> Any:
+    if timeout_ms is not None:
+        coro = asyncio.wait_for(coro, timeout=timeout_ms / 1000.0)
+
+    try:
+        if _GLOBAL_LOOP is not None and _GLOBAL_LOOP.is_running():
+            try:
+                curr_loop = asyncio.get_running_loop()
+                if curr_loop is _GLOBAL_LOOP:
+                    pass
+            except RuntimeError:
+                future = asyncio.run_coroutine_threadsafe(coro, _GLOBAL_LOOP)
+                return future.result(
+                    timeout=(timeout_ms / 1000.0) if timeout_ms is not None else None
+                )
+
+        try:
+            asyncio.get_running_loop()
+            return run_coro_in_new_loop(coro)
+        except RuntimeError:
+            loop = get_thread_loop()
+            return loop.run_until_complete(coro)
+    except asyncio.TimeoutError as exc:
+        raise TimeoutError(f"Isolation RPC timeout (timeout_ms={timeout_ms})") from exc
+    except concurrent.futures.TimeoutError as exc:
+        raise TimeoutError(f"Isolation RPC timeout (timeout_ms={timeout_ms})") from exc
+
+
+def call_singleton_rpc(
+    caller: Any,
+    method_name: str,
+    *args: Any,
+    timeout_ms: Optional[int] = None,
+    **kwargs: Any,
+) -> Any:
+    if caller is None:
+        raise RuntimeError(f"No RPC caller available for {method_name}")
+    method = getattr(caller, method_name)
+    return run_sync_rpc_coro(method(*args, **kwargs), timeout_ms=timeout_ms)
+
+
+class BaseProxy(Generic[T]):
+    _registry_class: type = BaseRegistry  # type: ignore[type-arg]
+    __module__: str = "comfy.isolation.proxies.base"
+    _TIMEOUT_RPC_METHODS = frozenset(
+        {
+            "partially_load",
+            "partially_unload",
+            "load",
+            "patch_model",
+            "unpatch_model",
+            "inner_model_apply_model",
+            "memory_required",
+            "model_dtype",
+            "inner_model_memory_required",
+            "inner_model_extra_conds_shapes",
+            "inner_model_extra_conds",
+            "process_latent_in",
+            "process_latent_out",
+            "scale_latent_inpaint",
+        }
+    )
+
+    def __init__(
+        self,
+        instance_id: str,
+        registry: Optional[Any] = None,
+        manage_lifecycle: bool = False,
+    ) -> None:
+        self._instance_id = instance_id
+        self._rpc_caller: Optional[Any] = None
+        self._registry = registry if registry is not None else self._registry_class()
+        self._manage_lifecycle = manage_lifecycle
+        self._cleaned_up = False
+        if manage_lifecycle and not IS_CHILD_PROCESS:
+            self._finalizer = weakref.finalize(
+                self, self._registry.unregister_sync, instance_id
+            )
+
+    def _get_rpc(self) -> Any:
+        if self._rpc_caller is None:
+            from pyisolate._internal.rpc_protocol import get_child_rpc_instance
+
+            rpc = get_child_rpc_instance()
+            if rpc is None:
+                raise RuntimeError(f"[{self.__class__.__name__}] No RPC in child")
+            self._rpc_caller = rpc.create_caller(
+                self._registry_class, self._registry_class.get_remote_id()
+            )
+        return self._rpc_caller
+
+    def _rpc_timeout_ms_for_method(self, method_name: str) -> Optional[int]:
+        if method_name not in self._TIMEOUT_RPC_METHODS:
+            return None
+        try:
+            timeout_ms = int(
+                os.environ.get("COMFY_ISOLATION_LOAD_RPC_TIMEOUT_MS", "120000")
+            )
+        except ValueError:
+            timeout_ms = 120000
+        return max(1, timeout_ms)
+
+    def _call_rpc(self, method_name: str, *args: Any, **kwargs: Any) -> Any:
+        rpc = self._get_rpc()
+        method = getattr(rpc, method_name)
+        timeout_ms = self._rpc_timeout_ms_for_method(method_name)
+        coro = method(self._instance_id, *args, **kwargs)
+        if timeout_ms is not None:
+            coro = asyncio.wait_for(coro, timeout=timeout_ms / 1000.0)
+
+        start_epoch = time.time()
+        start_perf = time.perf_counter()
+        thread_id = threading.get_ident()
+        try:
+            running_loop = asyncio.get_running_loop()
+            loop_id: Optional[int] = id(running_loop)
+        except RuntimeError:
+            loop_id = None
+        logger.debug(
+            "ISO:rpc_start proxy=%s method=%s instance_id=%s start_ts=%.6f "
+            "thread=%s loop=%s timeout_ms=%s",
+            self.__class__.__name__,
+            method_name,
+            self._instance_id,
+            start_epoch,
+            thread_id,
+            loop_id,
+            timeout_ms,
+        )
+
+        try:
+            return run_sync_rpc_coro(coro, timeout_ms=timeout_ms)
+        except TimeoutError as exc:
+            raise TimeoutError(
+                f"Isolation RPC timeout in {self.__class__.__name__}.{method_name} "
+                f"(instance_id={self._instance_id}, timeout_ms={timeout_ms})"
+            ) from exc
+        finally:
+            end_epoch = time.time()
+            elapsed_ms = (time.perf_counter() - start_perf) * 1000.0
+            logger.debug(
+                "ISO:rpc_end proxy=%s method=%s instance_id=%s end_ts=%.6f "
+                "elapsed_ms=%.3f thread=%s loop=%s",
+                self.__class__.__name__,
+                method_name,
+                self._instance_id,
+                end_epoch,
+                elapsed_ms,
+                thread_id,
+                loop_id,
+            )
+
+    def __getstate__(self) -> Dict[str, Any]:
+        return {"_instance_id": self._instance_id}
+
+    def __setstate__(self, state: Dict[str, Any]) -> None:
+        self._instance_id = state["_instance_id"]
+        self._rpc_caller = None
+        self._registry = self._registry_class()
+        self._manage_lifecycle = False
+        self._cleaned_up = False
+
+    def cleanup(self) -> None:
+        if self._cleaned_up or IS_CHILD_PROCESS:
+            return
+        self._cleaned_up = True
+        finalizer = getattr(self, "_finalizer", None)
+        if finalizer is not None:
+            finalizer.detach()
+        self._registry.unregister_sync(self._instance_id)
+
+    def __repr__(self) -> str:
+        return f"<{self.__class__.__name__} {self._instance_id}>"
+
+
+def create_rpc_method(method_name: str) -> Callable[..., Any]:
+    def method(self: BaseProxy[Any], *args: Any, **kwargs: Any) -> Any:
+        return self._call_rpc(method_name, *args, **kwargs)
+
+    method.__name__ = method_name
+    return method

comfy/isolation/proxies/folder_paths_proxy.py

@@ -0,0 +1,206 @@
+from __future__ import annotations
+import logging
+import os
+
+from typing import Any, Dict, Optional
+
+from pyisolate import ProxiedSingleton
+
+from .base import call_singleton_rpc
+
+_fp_logger = logging.getLogger(__name__)
+
+
+def _folder_paths():
+    import folder_paths
+
+    return folder_paths
+
+
+def _is_child_process() -> bool:
+    return os.environ.get("PYISOLATE_CHILD") == "1"
+
+
+def _serialize_folder_names_and_paths(data: dict[str, tuple[list[str], set[str]]]) -> dict[str, dict[str, list[str]]]:
+    return {
+        key: {"paths": list(paths), "extensions": sorted(list(extensions))}
+        for key, (paths, extensions) in data.items()
+    }
+
+
+def _deserialize_folder_names_and_paths(data: dict[str, dict[str, list[str]]]) -> dict[str, tuple[list[str], set[str]]]:
+    return {
+        key: (list(value.get("paths", [])), set(value.get("extensions", [])))
+        for key, value in data.items()
+    }
+
+
+class FolderPathsProxy(ProxiedSingleton):
+    """
+    Dynamic proxy for folder_paths.
+    Uses __getattr__ for most lookups, with explicit handling for
+    mutable collections to ensure efficient by-value transfer.
+    """
+
+    _rpc: Optional[Any] = None
+
+    @classmethod
+    def set_rpc(cls, rpc: Any) -> None:
+        cls._rpc = rpc.create_caller(cls, cls.get_remote_id())
+
+    @classmethod
+    def clear_rpc(cls) -> None:
+        cls._rpc = None
+
+    @classmethod
+    def _get_caller(cls) -> Any:
+        if cls._rpc is None:
+            raise RuntimeError("FolderPathsProxy RPC caller is not configured")
+        return cls._rpc
+
+    def __getattr__(self, name):
+        if _is_child_process():
+            property_rpc = {
+                "models_dir": "rpc_get_models_dir",
+                "folder_names_and_paths": "rpc_get_folder_names_and_paths",
+                "extension_mimetypes_cache": "rpc_get_extension_mimetypes_cache",
+                "filename_list_cache": "rpc_get_filename_list_cache",
+            }
+            rpc_name = property_rpc.get(name)
+            if rpc_name is not None:
+                return call_singleton_rpc(self._get_caller(), rpc_name)
+            raise AttributeError(name)
+        return getattr(_folder_paths(), name)
+
+    @property
+    def folder_names_and_paths(self) -> Dict:
+        if _is_child_process():
+            payload = call_singleton_rpc(self._get_caller(), "rpc_get_folder_names_and_paths")
+            return _deserialize_folder_names_and_paths(payload)
+        return _folder_paths().folder_names_and_paths
+
+    @property
+    def extension_mimetypes_cache(self) -> Dict:
+        if _is_child_process():
+            return dict(call_singleton_rpc(self._get_caller(), "rpc_get_extension_mimetypes_cache"))
+        return dict(_folder_paths().extension_mimetypes_cache)
+
+    @property
+    def filename_list_cache(self) -> Dict:
+        if _is_child_process():
+            return dict(call_singleton_rpc(self._get_caller(), "rpc_get_filename_list_cache"))
+        return dict(_folder_paths().filename_list_cache)
+
+    @property
+    def models_dir(self) -> str:
+        if _is_child_process():
+            return str(call_singleton_rpc(self._get_caller(), "rpc_get_models_dir"))
+        return _folder_paths().models_dir
+
+    def get_temp_directory(self) -> str:
+        if _is_child_process():
+            return call_singleton_rpc(self._get_caller(), "rpc_get_temp_directory")
+        return _folder_paths().get_temp_directory()
+
+    def get_input_directory(self) -> str:
+        if _is_child_process():
+            return call_singleton_rpc(self._get_caller(), "rpc_get_input_directory")
+        return _folder_paths().get_input_directory()
+
+    def get_output_directory(self) -> str:
+        if _is_child_process():
+            return call_singleton_rpc(self._get_caller(), "rpc_get_output_directory")
+        return _folder_paths().get_output_directory()
+
+    def get_user_directory(self) -> str:
+        if _is_child_process():
+            return call_singleton_rpc(self._get_caller(), "rpc_get_user_directory")
+        return _folder_paths().get_user_directory()
+
+    def get_annotated_filepath(self, name: str, default_dir: str | None = None) -> str:
+        if _is_child_process():
+            return call_singleton_rpc(
+                self._get_caller(), "rpc_get_annotated_filepath", name, default_dir
+            )
+        return _folder_paths().get_annotated_filepath(name, default_dir)
+
+    def exists_annotated_filepath(self, name: str) -> bool:
+        if _is_child_process():
+            return bool(
+                call_singleton_rpc(self._get_caller(), "rpc_exists_annotated_filepath", name)
+            )
+        return bool(_folder_paths().exists_annotated_filepath(name))
+
+    def add_model_folder_path(
+        self, folder_name: str, full_folder_path: str, is_default: bool = False
+    ) -> None:
+        if _is_child_process():
+            call_singleton_rpc(
+                self._get_caller(),
+                "rpc_add_model_folder_path",
+                folder_name,
+                full_folder_path,
+                is_default,
+            )
+            return None
+        _folder_paths().add_model_folder_path(folder_name, full_folder_path, is_default)
+        return None
+
+    def get_folder_paths(self, folder_name: str) -> list[str]:
+        if _is_child_process():
+            return list(call_singleton_rpc(self._get_caller(), "rpc_get_folder_paths", folder_name))
+        return list(_folder_paths().get_folder_paths(folder_name))
+
+    def get_filename_list(self, folder_name: str) -> list[str]:
+        if _is_child_process():
+            return list(call_singleton_rpc(self._get_caller(), "rpc_get_filename_list", folder_name))
+        return list(_folder_paths().get_filename_list(folder_name))
+
+    def get_full_path(self, folder_name: str, filename: str) -> str | None:
+        if _is_child_process():
+            return call_singleton_rpc(self._get_caller(), "rpc_get_full_path", folder_name, filename)
+        return _folder_paths().get_full_path(folder_name, filename)
+
+    async def rpc_get_models_dir(self) -> str:
+        return _folder_paths().models_dir
+
+    async def rpc_get_folder_names_and_paths(self) -> dict[str, dict[str, list[str]]]:
+        return _serialize_folder_names_and_paths(_folder_paths().folder_names_and_paths)
+
+    async def rpc_get_extension_mimetypes_cache(self) -> dict[str, Any]:
+        return dict(_folder_paths().extension_mimetypes_cache)
+
+    async def rpc_get_filename_list_cache(self) -> dict[str, Any]:
+        return dict(_folder_paths().filename_list_cache)
+
+    async def rpc_get_temp_directory(self) -> str:
+        return _folder_paths().get_temp_directory()
+
+    async def rpc_get_input_directory(self) -> str:
+        return _folder_paths().get_input_directory()
+
+    async def rpc_get_output_directory(self) -> str:
+        return _folder_paths().get_output_directory()
+
+    async def rpc_get_user_directory(self) -> str:
+        return _folder_paths().get_user_directory()
+
+    async def rpc_get_annotated_filepath(self, name: str, default_dir: str | None = None) -> str:
+        return _folder_paths().get_annotated_filepath(name, default_dir)
+
+    async def rpc_exists_annotated_filepath(self, name: str) -> bool:
+        return _folder_paths().exists_annotated_filepath(name)
+
+    async def rpc_add_model_folder_path(
+        self, folder_name: str, full_folder_path: str, is_default: bool = False
+    ) -> None:
+        _folder_paths().add_model_folder_path(folder_name, full_folder_path, is_default)
+
+    async def rpc_get_folder_paths(self, folder_name: str) -> list[str]:
+        return _folder_paths().get_folder_paths(folder_name)
+
+    async def rpc_get_filename_list(self, folder_name: str) -> list[str]:
+        return _folder_paths().get_filename_list(folder_name)
+
+    async def rpc_get_full_path(self, folder_name: str, filename: str) -> str | None:
+        return _folder_paths().get_full_path(folder_name, filename)

comfy/isolation/proxies/helper_proxies.py

@@ -0,0 +1,158 @@
+from __future__ import annotations
+
+import os
+from typing import Any, Dict, Optional
+
+from pyisolate import ProxiedSingleton
+
+from .base import call_singleton_rpc
+
+
+class AnyTypeProxy(str):
+    """Replacement for custom AnyType objects used by some nodes."""
+
+    def __new__(cls, value: str = "*"):
+        return super().__new__(cls, value)
+
+    def __ne__(self, other):  # type: ignore[override]
+        return False
+
+
+class FlexibleOptionalInputProxy(dict):
+    """Replacement for FlexibleOptionalInputType to allow dynamic inputs."""
+
+    def __init__(self, flex_type, data: Optional[Dict[str, object]] = None):
+        super().__init__()
+        self.type = flex_type
+        if data:
+            self.update(data)
+
+    def __getitem__(self, key):  # type: ignore[override]
+        return (self.type,)
+
+    def __contains__(self, key):  # type: ignore[override]
+        return True
+
+
+class ByPassTypeTupleProxy(tuple):
+    """Replacement for ByPassTypeTuple to mirror wildcard fallback behavior."""
+
+    def __new__(cls, values):
+        return super().__new__(cls, values)
+
+    def __getitem__(self, index):  # type: ignore[override]
+        if index >= len(self):
+            return AnyTypeProxy("*")
+        return super().__getitem__(index)
+
+
+def _restore_special_value(value: Any) -> Any:
+    if isinstance(value, dict):
+        if value.get("__pyisolate_any_type__"):
+            return AnyTypeProxy(value.get("value", "*"))
+        if value.get("__pyisolate_flexible_optional__"):
+            flex_type = _restore_special_value(value.get("type"))
+            data_raw = value.get("data")
+            data = (
+                {k: _restore_special_value(v) for k, v in data_raw.items()}
+                if isinstance(data_raw, dict)
+                else {}
+            )
+            return FlexibleOptionalInputProxy(flex_type, data)
+        if value.get("__pyisolate_tuple__") is not None:
+            return tuple(
+                _restore_special_value(v) for v in value["__pyisolate_tuple__"]
+            )
+        if value.get("__pyisolate_bypass_tuple__") is not None:
+            return ByPassTypeTupleProxy(
+                tuple(
+                    _restore_special_value(v)
+                    for v in value["__pyisolate_bypass_tuple__"]
+                )
+            )
+        return {k: _restore_special_value(v) for k, v in value.items()}
+    if isinstance(value, list):
+        return [_restore_special_value(v) for v in value]
+    return value
+
+
+def _serialize_special_value(value: Any) -> Any:
+    if isinstance(value, AnyTypeProxy):
+        return {"__pyisolate_any_type__": True, "value": str(value)}
+    if isinstance(value, FlexibleOptionalInputProxy):
+        return {
+            "__pyisolate_flexible_optional__": True,
+            "type": _serialize_special_value(value.type),
+            "data": {k: _serialize_special_value(v) for k, v in value.items()},
+        }
+    if isinstance(value, ByPassTypeTupleProxy):
+        return {
+            "__pyisolate_bypass_tuple__": [_serialize_special_value(v) for v in value]
+        }
+    if isinstance(value, tuple):
+        return {"__pyisolate_tuple__": [_serialize_special_value(v) for v in value]}
+    if isinstance(value, list):
+        return [_serialize_special_value(v) for v in value]
+    if isinstance(value, dict):
+        return {k: _serialize_special_value(v) for k, v in value.items()}
+    return value
+
+
+def _restore_input_types_local(raw: Dict[str, object]) -> Dict[str, object]:
+    if not isinstance(raw, dict):
+        return raw  # type: ignore[return-value]
+
+    restored: Dict[str, object] = {}
+    for section, entries in raw.items():
+        if isinstance(entries, dict) and entries.get("__pyisolate_flexible_optional__"):
+            restored[section] = _restore_special_value(entries)
+        elif isinstance(entries, dict):
+            restored[section] = {
+                k: _restore_special_value(v) for k, v in entries.items()
+            }
+        else:
+            restored[section] = _restore_special_value(entries)
+    return restored
+
+
+class HelperProxiesService(ProxiedSingleton):
+    _rpc: Optional[Any] = None
+
+    @classmethod
+    def set_rpc(cls, rpc: Any) -> None:
+        cls._rpc = rpc.create_caller(cls, cls.get_remote_id())
+
+    @classmethod
+    def clear_rpc(cls) -> None:
+        cls._rpc = None
+
+    @classmethod
+    def _get_caller(cls) -> Any:
+        if cls._rpc is None:
+            raise RuntimeError("HelperProxiesService RPC caller is not configured")
+        return cls._rpc
+
+    async def rpc_restore_input_types(self, raw: Dict[str, object]) -> Dict[str, object]:
+        restored = _restore_input_types_local(raw)
+        return _serialize_special_value(restored)
+
+
+def restore_input_types(raw: Dict[str, object]) -> Dict[str, object]:
+    """Restore serialized INPUT_TYPES payload back into ComfyUI-compatible objects."""
+    if os.environ.get("PYISOLATE_CHILD") == "1":
+        payload = call_singleton_rpc(
+            HelperProxiesService._get_caller(),
+            "rpc_restore_input_types",
+            raw,
+        )
+        return _restore_input_types_local(payload)
+    return _restore_input_types_local(raw)
+
+
+__all__ = [
+    "AnyTypeProxy",
+    "FlexibleOptionalInputProxy",
+    "ByPassTypeTupleProxy",
+    "HelperProxiesService",
+    "restore_input_types",
+]

comfy/isolation/proxies/model_management_proxy.py

@@ -0,0 +1,142 @@
+from __future__ import annotations
+
+import os
+from typing import Any, Optional
+
+from pyisolate import ProxiedSingleton
+
+from .base import call_singleton_rpc
+
+
+def _mm():
+    import comfy.model_management
+
+    return comfy.model_management
+
+
+def _is_child_process() -> bool:
+    return os.environ.get("PYISOLATE_CHILD") == "1"
+
+
+class TorchDeviceProxy:
+    def __init__(self, device_str: str):
+        self._device_str = device_str
+        if ":" in device_str:
+            device_type, index = device_str.split(":", 1)
+            self.type = device_type
+            self.index = int(index)
+        else:
+            self.type = device_str
+            self.index = None
+
+    def __str__(self) -> str:
+        return self._device_str
+
+    def __repr__(self) -> str:
+        return f"TorchDeviceProxy({self._device_str!r})"
+
+
+def _serialize_value(value: Any) -> Any:
+    value_type = type(value)
+    if value_type.__module__ == "torch" and value_type.__name__ == "device":
+        return {"__pyisolate_torch_device__": str(value)}
+    if isinstance(value, TorchDeviceProxy):
+        return {"__pyisolate_torch_device__": str(value)}
+    if isinstance(value, tuple):
+        return {"__pyisolate_tuple__": [_serialize_value(item) for item in value]}
+    if isinstance(value, list):
+        return [_serialize_value(item) for item in value]
+    if isinstance(value, dict):
+        return {key: _serialize_value(inner) for key, inner in value.items()}
+    return value
+
+
+def _deserialize_value(value: Any) -> Any:
+    if isinstance(value, dict):
+        if "__pyisolate_torch_device__" in value:
+            return TorchDeviceProxy(value["__pyisolate_torch_device__"])
+        if "__pyisolate_tuple__" in value:
+            return tuple(_deserialize_value(item) for item in value["__pyisolate_tuple__"])
+        return {key: _deserialize_value(inner) for key, inner in value.items()}
+    if isinstance(value, list):
+        return [_deserialize_value(item) for item in value]
+    return value
+
+
+def _normalize_argument(value: Any) -> Any:
+    if isinstance(value, TorchDeviceProxy):
+        import torch
+
+        return torch.device(str(value))
+    if isinstance(value, dict):
+        if "__pyisolate_torch_device__" in value:
+            import torch
+
+            return torch.device(value["__pyisolate_torch_device__"])
+        if "__pyisolate_tuple__" in value:
+            return tuple(_normalize_argument(item) for item in value["__pyisolate_tuple__"])
+        return {key: _normalize_argument(inner) for key, inner in value.items()}
+    if isinstance(value, list):
+        return [_normalize_argument(item) for item in value]
+    return value
+
+
+class ModelManagementProxy(ProxiedSingleton):
+    """
+    Exact-relay proxy for comfy.model_management.
+    Child calls never import comfy.model_management directly; they serialize
+    arguments, relay to host, and deserialize the host result back.
+    """
+
+    _rpc: Optional[Any] = None
+
+    @classmethod
+    def set_rpc(cls, rpc: Any) -> None:
+        cls._rpc = rpc.create_caller(cls, cls.get_remote_id())
+
+    @classmethod
+    def clear_rpc(cls) -> None:
+        cls._rpc = None
+
+    @classmethod
+    def _get_caller(cls) -> Any:
+        if cls._rpc is None:
+            raise RuntimeError("ModelManagementProxy RPC caller is not configured")
+        return cls._rpc
+
+    def _relay_call(self, method_name: str, *args: Any, **kwargs: Any) -> Any:
+        payload = call_singleton_rpc(
+            self._get_caller(),
+            "rpc_call",
+            method_name,
+            _serialize_value(args),
+            _serialize_value(kwargs),
+        )
+        return _deserialize_value(payload)
+
+    @property
+    def VRAMState(self):
+        return _mm().VRAMState
+
+    @property
+    def CPUState(self):
+        return _mm().CPUState
+
+    @property
+    def OOM_EXCEPTION(self):
+        return _mm().OOM_EXCEPTION
+
+    def __getattr__(self, name: str):
+        if _is_child_process():
+            def child_method(*args: Any, **kwargs: Any) -> Any:
+                return self._relay_call(name, *args, **kwargs)
+
+            return child_method
+        return getattr(_mm(), name)
+
+    async def rpc_call(self, method_name: str, args: Any, kwargs: Any) -> Any:
+        normalized_args = _normalize_argument(_deserialize_value(args))
+        normalized_kwargs = _normalize_argument(_deserialize_value(kwargs))
+        method = getattr(_mm(), method_name)
+        result = method(*normalized_args, **normalized_kwargs)
+        return _serialize_value(result)

comfy/isolation/proxies/progress_proxy.py

@@ -0,0 +1,87 @@
+from __future__ import annotations
+
+import logging
+import os
+from typing import Any, Optional
+
+try:
+    from pyisolate import ProxiedSingleton
+except ImportError:
+
+    class ProxiedSingleton:
+        pass
+
+from .base import call_singleton_rpc
+
+
+def _get_progress_state():
+    from comfy_execution.progress import get_progress_state
+
+    return get_progress_state()
+
+
+def _is_child_process() -> bool:
+    return os.environ.get("PYISOLATE_CHILD") == "1"
+
+logger = logging.getLogger(__name__)
+
+
+class ProgressProxy(ProxiedSingleton):
+    _rpc: Optional[Any] = None
+
+    @classmethod
+    def set_rpc(cls, rpc: Any) -> None:
+        cls._rpc = rpc.create_caller(cls, cls.get_remote_id())
+
+    @classmethod
+    def clear_rpc(cls) -> None:
+        cls._rpc = None
+
+    @classmethod
+    def _get_caller(cls) -> Any:
+        if cls._rpc is None:
+            raise RuntimeError("ProgressProxy RPC caller is not configured")
+        return cls._rpc
+
+    def set_progress(
+        self,
+        value: float,
+        max_value: float,
+        node_id: Optional[str] = None,
+        image: Any = None,
+    ) -> None:
+        if _is_child_process():
+            call_singleton_rpc(
+                self._get_caller(),
+                "rpc_set_progress",
+                value,
+                max_value,
+                node_id,
+                image,
+            )
+            return None
+
+        _get_progress_state().update_progress(
+            node_id=node_id,
+            value=value,
+            max_value=max_value,
+            image=image,
+        )
+        return None
+
+    async def rpc_set_progress(
+        self,
+        value: float,
+        max_value: float,
+        node_id: Optional[str] = None,
+        image: Any = None,
+    ) -> None:
+        _get_progress_state().update_progress(
+            node_id=node_id,
+            value=value,
+            max_value=max_value,
+            image=image,
+        )
+
+
+__all__ = ["ProgressProxy"]

comfy/isolation/proxies/prompt_server_impl.py

@@ -0,0 +1,306 @@
+# pylint: disable=import-outside-toplevel,logging-fstring-interpolation,redefined-outer-name,reimported,super-init-not-called
+"""Stateless RPC Implementation for PromptServer.
+
+Replaces the legacy PromptServerProxy (Singleton) with a clean Service/Stub architecture.
+- Host: PromptServerService (RPC Handler)
+- Child: PromptServerStub (Interface Implementation)
+"""
+
+from __future__ import annotations
+
+import os
+from typing import Any, Dict, Optional, Callable
+
+import logging
+
+# IMPORTS
+from pyisolate import ProxiedSingleton
+from .base import call_singleton_rpc
+
+logger = logging.getLogger(__name__)
+LOG_PREFIX = "[Isolation:C<->H]"
+
+# ...
+
+# =============================================================================
+# CHILD SIDE: PromptServerStub
+# =============================================================================
+
+
+class PromptServerStub:
+    """Stateless Stub for PromptServer."""
+
+    # Masquerade as the real server module
+    __module__ = "server"
+
+    _instance: Optional["PromptServerStub"] = None
+    _rpc: Optional[Any] = None  # This will be the Caller object
+    _source_file: Optional[str] = None
+
+    def __init__(self):
+        self.routes = RouteStub(self)
+
+    @classmethod
+    def set_rpc(cls, rpc: Any) -> None:
+        """Inject RPC client (called by adapter.py or manually)."""
+        # Create caller for HOST Service
+        # Assuming Host Service is registered as "PromptServerService" (class name)
+        # We target the Host Service Class
+        target_id = "PromptServerService"
+        # We need to pass a class to create_caller? Usually yes.
+        # But we don't have the Service class imported here necessarily (if running on child).
+        # pyisolate check verify_service type?
+        # If we pass PromptServerStub as the 'class', it might mismatch if checking types.
+        # But we can try passing PromptServerStub if it mirrors the service name? No, stub is PromptServerStub.
+        # We need a dummy class with right name?
+        # Or just rely on string ID if create_caller supports it?
+        # Standard: rpc.create_caller(PromptServerStub, target_id)
+        # But wait, PromptServerStub is the *Local* class.
+        # We want to call *Remote* class.
+        # If we use PromptServerStub as the type, returning object will be typed as PromptServerStub?
+        # The first arg is 'service_cls'.
+        cls._rpc = rpc.create_caller(
+            PromptServerService, target_id
+        )  # We import Service below?
+
+    @classmethod
+    def clear_rpc(cls) -> None:
+        cls._rpc = None
+
+    # We need PromptServerService available for the create_caller call?
+    # Or just use the Stub class if ID matches?
+    # prompt_server_impl.py defines BOTH. So PromptServerService IS available!
+
+    @property
+    def instance(self) -> "PromptServerStub":
+        return self
+
+    # ... Compatibility ...
+    @classmethod
+    def _get_source_file(cls) -> str:
+        if cls._source_file is None:
+            import folder_paths
+
+            cls._source_file = os.path.join(folder_paths.base_path, "server.py")
+        return cls._source_file
+
+    @property
+    def __file__(self) -> str:
+        return self._get_source_file()
+
+    # --- Properties ---
+    @property
+    def client_id(self) -> Optional[str]:
+        return "isolated_client"
+
+    @property
+    def supports(self) -> set:
+        return {"custom_nodes_from_web"}
+
+    @property
+    def app(self):
+        return _AppStub(self)
+
+    @property
+    def prompt_queue(self):
+        raise RuntimeError(
+            "PromptServer.prompt_queue is not accessible in isolated nodes."
+        )
+
+    # --- UI Communication (RPC Delegates) ---
+    async def send_sync(
+        self, event: str, data: Dict[str, Any], sid: Optional[str] = None
+    ) -> None:
+        if self._rpc:
+            await self._rpc.ui_send_sync(event, data, sid)
+
+    async def send(
+        self, event: str, data: Dict[str, Any], sid: Optional[str] = None
+    ) -> None:
+        if self._rpc:
+            await self._rpc.ui_send(event, data, sid)
+
+    def send_progress_text(self, text: str, node_id: str, sid=None) -> None:
+        if self._rpc:
+            # Fire and forget likely needed. If method is async on host, caller invocation returns coroutine.
+            # We must schedule it?
+            # Or use fire_remote equivalent?
+            # Caller object usually proxies calls. If host method is async, it returns coro.
+            # If we are sync here (send_progress_text checks imply sync usage), we must background it.
+            # But UtilsProxy hook wrapper creates task.
+            # Does send_progress_text need to be sync? Yes, node code calls it sync.
+            import asyncio
+
+            try:
+                loop = asyncio.get_running_loop()
+                loop.create_task(self._rpc.ui_send_progress_text(text, node_id, sid))
+            except RuntimeError:
+                call_singleton_rpc(self._rpc, "ui_send_progress_text", text, node_id, sid)
+
+    # --- Route Registration Logic ---
+    _pending_child_routes: list = []
+
+    def register_route(self, method: str, path: str, handler: Callable):
+        """Buffer route registration. Routes are flushed via flush_child_routes()."""
+        PromptServerStub._pending_child_routes.append((method, path, handler))
+        logger.info("%s Buffered isolated route %s %s", LOG_PREFIX, method, path)
+
+    @classmethod
+    async def flush_child_routes(cls):
+        """Send all buffered route registrations to host via RPC. Call from on_module_loaded."""
+        if not cls._rpc:
+            return 0
+        flushed = 0
+        for method, path, handler in cls._pending_child_routes:
+            try:
+                await cls._rpc.register_route_rpc(method, path, handler)
+                flushed += 1
+            except Exception as e:
+                logger.error("%s Child route flush failed %s %s: %s", LOG_PREFIX, method, path, e)
+        cls._pending_child_routes = []
+        return flushed
+
+
+class RouteStub:
+    """Simulates aiohttp.web.RouteTableDef."""
+
+    def __init__(self, stub: PromptServerStub):
+        self._stub = stub
+
+    def get(self, path: str):
+        def decorator(handler):
+            self._stub.register_route("GET", path, handler)
+            return handler
+
+        return decorator
+
+    def post(self, path: str):
+        def decorator(handler):
+            self._stub.register_route("POST", path, handler)
+            return handler
+
+        return decorator
+
+    def patch(self, path: str):
+        def decorator(handler):
+            self._stub.register_route("PATCH", path, handler)
+            return handler
+
+        return decorator
+
+    def put(self, path: str):
+        def decorator(handler):
+            self._stub.register_route("PUT", path, handler)
+            return handler
+
+        return decorator
+
+    def delete(self, path: str):
+        def decorator(handler):
+            self._stub.register_route("DELETE", path, handler)
+            return handler
+
+        return decorator
+
+
+# =============================================================================
+# HOST SIDE: PromptServerService
+# =============================================================================
+
+
+class PromptServerService(ProxiedSingleton):
+    """Host-side RPC Service for PromptServer."""
+
+    def __init__(self):
+        pass
+
+    @property
+    def server(self):
+        from server import PromptServer
+
+        return PromptServer.instance
+
+    async def ui_send_sync(
+        self, event: str, data: Dict[str, Any], sid: Optional[str] = None
+    ):
+        await self.server.send_sync(event, data, sid)
+
+    async def ui_send(
+        self, event: str, data: Dict[str, Any], sid: Optional[str] = None
+    ):
+        await self.server.send(event, data, sid)
+
+    async def ui_send_progress_text(self, text: str, node_id: str, sid=None):
+        # Made async to be awaitable by RPC layer
+        self.server.send_progress_text(text, node_id, sid)
+
+    async def register_route_rpc(self, method: str, path: str, child_handler_proxy):
+        """RPC Target: Register a route that forwards to the Child."""
+        from aiohttp import web
+        logger.info("%s Registering isolated route %s %s", LOG_PREFIX, method, path)
+
+        async def route_wrapper(request: web.Request) -> web.Response:
+            # 1. Capture request data
+            req_data = {
+                "method": request.method,
+                "path": request.path,
+                "query": dict(request.query),
+            }
+            if request.can_read_body:
+                req_data["text"] = await request.text()
+
+            try:
+                # 2. Call Child Handler via RPC (child_handler_proxy is async callable)
+                result = await child_handler_proxy(req_data)
+
+                # 3. Serialize Response
+                return self._serialize_response(result)
+            except Exception as e:
+                logger.error(f"{LOG_PREFIX} Isolated Route Error: {e}")
+                return web.Response(status=500, text=str(e))
+
+        self.server.app.router.add_route(method, path, route_wrapper)
+        logger.info("%s Registered isolated route %s %s", LOG_PREFIX, method, path)
+
+    def _serialize_response(self, result: Any) -> Any:
+        """Helper to convert Child result -> web.Response"""
+        from aiohttp import web
+        if isinstance(result, web.Response):
+            return result
+        # Handle dict (json)
+        if isinstance(result, dict):
+            return web.json_response(result)
+        # Handle string
+        if isinstance(result, str):
+            return web.Response(text=result)
+        # Fallback
+        return web.Response(text=str(result))
+
+
+class _RouterStub:
+    """Captures router.add_route and router.add_static calls in isolation child."""
+
+    def __init__(self, stub):
+        self._stub = stub
+
+    def add_route(self, method, path, handler, **kwargs):
+        self._stub.register_route(method, path, handler)
+
+    def add_static(self, prefix, path, **kwargs):
+        # Static file serving not supported in isolation — silently skip
+        pass
+
+
+class _AppStub:
+    """Captures PromptServer.app access patterns in isolation child."""
+
+    def __init__(self, stub):
+        self.router = _RouterStub(stub)
+        self.frozen = False
+
+    def add_routes(self, routes):
+        # aiohttp route table — iterate and register each
+        for route in routes:
+            if hasattr(route, "method") and hasattr(route, "handler"):
+                self.router.add_route(route.method, route.path, route.handler)
+            # StaticDef and other non-method routes — silently skip

comfy/isolation/proxies/utils_proxy.py

@@ -0,0 +1,64 @@
+# pylint: disable=cyclic-import,import-outside-toplevel
+from __future__ import annotations
+
+from typing import Optional, Any
+from pyisolate import ProxiedSingleton
+
+import os
+
+
+def _comfy_utils():
+    import comfy.utils
+    return comfy.utils
+
+
+class UtilsProxy(ProxiedSingleton):
+    """
+    Proxy for comfy.utils.
+    Primarily handles the PROGRESS_BAR_HOOK to ensure progress updates
+    from isolated nodes reach the host.
+    """
+
+    # _instance and __new__ removed to rely on SingletonMetaclass
+    _rpc: Optional[Any] = None
+
+    @classmethod
+    def set_rpc(cls, rpc: Any) -> None:
+        # Create caller using class name as ID (standard for Singletons)
+        cls._rpc = rpc.create_caller(cls, "UtilsProxy")
+
+    @classmethod
+    def clear_rpc(cls) -> None:
+        cls._rpc = None
+
+    async def progress_bar_hook(
+        self,
+        value: int,
+        total: int,
+        preview: Optional[bytes] = None,
+        node_id: Optional[str] = None,
+    ) -> Any:
+        """
+        Host-side implementation: forwards the call to the real global hook.
+        Child-side: this method call is intercepted by RPC and sent to host.
+        """
+        if os.environ.get("PYISOLATE_CHILD") == "1":
+            if UtilsProxy._rpc is None:
+                raise RuntimeError("UtilsProxy RPC caller is not configured")
+            return await UtilsProxy._rpc.progress_bar_hook(
+                value, total, preview, node_id
+            )
+
+        # Host Execution
+        utils = _comfy_utils()
+        if utils.PROGRESS_BAR_HOOK is not None:
+            return utils.PROGRESS_BAR_HOOK(value, total, preview, node_id)
+        return None
+
+    def set_progress_bar_global_hook(self, hook: Any) -> None:
+        """Forward hook registration (though usually not needed from child)."""
+        if os.environ.get("PYISOLATE_CHILD") == "1":
+            raise RuntimeError(
+                "UtilsProxy.set_progress_bar_global_hook is not available in child without exact relay support"
+            )
+        _comfy_utils().set_progress_bar_global_hook(hook)

comfy/isolation/proxies/web_directory_proxy.py

@@ -0,0 +1,229 @@
+"""WebDirectoryProxy — serves isolated node web assets via RPC.
+
+Child side: enumerates and reads files from the extension's web/ directory.
+Host side: gets an RPC proxy that fetches file listings and contents on demand.
+
+Only files with allowed extensions (.js, .html, .css) are served.
+Directory traversal is rejected. File contents are base64-encoded for
+safe JSON-RPC transport.
+"""
+
+from __future__ import annotations
+
+import base64
+import binascii
+import logging
+import os
+from pathlib import Path, PurePosixPath
+from typing import Any, Dict, List
+
+from pyisolate import ProxiedSingleton
+
+logger = logging.getLogger(__name__)
+
+ALLOWED_EXTENSIONS = frozenset({".js", ".html", ".css"})
+
+MIME_TYPES = {
+    ".js": "application/javascript",
+    ".html": "text/html",
+    ".css": "text/css",
+}
+
+
+class WebDirectoryProxy(ProxiedSingleton):
+    """Proxy for serving isolated extension web directories.
+
+    On the child side, this class has direct filesystem access to the
+    extension's web/ directory.  On the host side, callers get an RPC
+    proxy whose method calls are forwarded to the child.
+    """
+
+    # {extension_name: absolute_path_to_web_dir}
+    _web_dirs: dict[str, str] = {}
+
+    @classmethod
+    def register_web_dir(cls, extension_name: str, web_dir_path: str) -> None:
+        """Register an extension's web directory (child-side only)."""
+        cls._web_dirs[extension_name] = web_dir_path
+        logger.info(
+            "][ WebDirectoryProxy: registered %s -> %s",
+            extension_name,
+            web_dir_path,
+        )
+
+    def list_web_files(self, extension_name: str) -> List[Dict[str, str]]:
+        """Return a list of servable files in the extension's web directory.
+
+        Each entry is {"relative_path": "js/foo.js", "content_type": "application/javascript"}.
+        Only files with allowed extensions are included.
+        """
+        web_dir = self._web_dirs.get(extension_name)
+        if not web_dir:
+            return []
+
+        root = Path(web_dir)
+        if not root.is_dir():
+            return []
+
+        result: List[Dict[str, str]] = []
+        for path in sorted(root.rglob("*")):
+            if not path.is_file():
+                continue
+            ext = path.suffix.lower()
+            if ext not in ALLOWED_EXTENSIONS:
+                continue
+            rel = path.relative_to(root)
+            result.append({
+                "relative_path": str(PurePosixPath(rel)),
+                "content_type": MIME_TYPES[ext],
+            })
+        return result
+
+    def get_web_file(
+        self, extension_name: str, relative_path: str
+    ) -> Dict[str, Any]:
+        """Return the contents of a single web file as base64.
+
+        Raises ValueError for traversal attempts or disallowed file types.
+        Returns {"content": <base64 str>, "content_type": <MIME str>}.
+        """
+        _validate_path(relative_path)
+
+        web_dir = self._web_dirs.get(extension_name)
+        if not web_dir:
+            raise FileNotFoundError(
+                f"No web directory registered for {extension_name}"
+            )
+
+        root = Path(web_dir).resolve()
+        target = (root / relative_path).resolve()
+
+        # Ensure resolved path is under the web directory
+        if os.path.commonpath([str(root), str(target)]) != str(root):
+            raise ValueError(f"Path escapes web directory: {relative_path}")
+
+        if not target.is_file():
+            raise FileNotFoundError(f"File not found: {relative_path}")
+
+        ext = target.suffix.lower()
+        if ext not in ALLOWED_EXTENSIONS:
+            raise ValueError(f"Disallowed file type: {ext}")
+
+        content_type = MIME_TYPES[ext]
+        raw = target.read_bytes()
+
+        return {
+            "content": base64.b64encode(raw).decode("ascii"),
+            "content_type": content_type,
+        }
+
+
+def _validate_path(relative_path: str) -> None:
+    """Reject directory traversal and absolute paths."""
+    if os.path.isabs(relative_path):
+        raise ValueError(f"Absolute paths are not allowed: {relative_path}")
+    if ".." in PurePosixPath(relative_path).parts:
+        raise ValueError(f"Directory traversal is not allowed: {relative_path}")
+
+
+# ---------------------------------------------------------------------------
+# Host-side cache and aiohttp handler
+# ---------------------------------------------------------------------------
+
+
+class WebDirectoryCache:
+    """Host-side in-memory cache for proxied web directory contents.
+
+    Populated lazily via RPC calls to the child's WebDirectoryProxy.
+    Once a file is cached, subsequent requests are served from memory.
+    """
+
+    def __init__(self) -> None:
+        # {extension_name: {relative_path: {"content": bytes, "content_type": str}}}
+        self._file_cache: dict[str, dict[str, dict[str, Any]]] = {}
+        # {extension_name: [{"relative_path": str, "content_type": str}, ...]}
+        self._listing_cache: dict[str, list[dict[str, str]]] = {}
+        # {extension_name: WebDirectoryProxy (RPC proxy instance)}
+        self._proxies: dict[str, Any] = {}
+
+    def register_proxy(self, extension_name: str, proxy: Any) -> None:
+        """Register an RPC proxy for an extension's web directory."""
+        self._proxies[extension_name] = proxy
+        logger.info(
+            "][ WebDirectoryCache: registered proxy for %s", extension_name
+        )
+
+    @property
+    def extension_names(self) -> list[str]:
+        return list(self._proxies.keys())
+
+    def list_files(self, extension_name: str) -> list[dict[str, str]]:
+        """List servable files for an extension (cached after first call)."""
+        if extension_name not in self._listing_cache:
+            proxy = self._proxies.get(extension_name)
+            if proxy is None:
+                return []
+            try:
+                self._listing_cache[extension_name] = proxy.list_web_files(
+                    extension_name
+                )
+            except Exception:
+                logger.warning(
+                    "][ WebDirectoryCache: failed to list files for %s",
+                    extension_name,
+                    exc_info=True,
+                )
+                return []
+        return self._listing_cache[extension_name]
+
+    def get_file(
+        self, extension_name: str, relative_path: str
+    ) -> dict[str, Any] | None:
+        """Get file content (cached after first fetch). Returns None on miss."""
+        ext_cache = self._file_cache.get(extension_name)
+        if ext_cache and relative_path in ext_cache:
+            return ext_cache[relative_path]
+
+        proxy = self._proxies.get(extension_name)
+        if proxy is None:
+            return None
+
+        try:
+            result = proxy.get_web_file(extension_name, relative_path)
+        except (FileNotFoundError, ValueError):
+            return None
+        except Exception:
+            logger.warning(
+                "][ WebDirectoryCache: failed to fetch %s/%s",
+                extension_name,
+                relative_path,
+                exc_info=True,
+            )
+            return None
+
+        try:
+            decoded = {
+                "content": base64.b64decode(result["content"], validate=True),
+                "content_type": result["content_type"],
+            }
+        except (binascii.Error, KeyError, TypeError):
+            logger.warning(
+                "][ WebDirectoryCache: invalid payload for %s/%s",
+                extension_name,
+                relative_path,
+                exc_info=True,
+            )
+            return None
+
+        if extension_name not in self._file_cache:
+            self._file_cache[extension_name] = {}
+        self._file_cache[extension_name][relative_path] = decoded
+        return decoded
+
+
+# Global cache instance — populated during isolation loading
+_web_directory_cache = WebDirectoryCache()
+
+
+def get_web_directory_cache() -> WebDirectoryCache:
+    return _web_directory_cache

comfy/isolation/rpc_bridge.py

@@ -0,0 +1,49 @@
+import asyncio
+import logging
+import threading
+
+logger = logging.getLogger(__name__)
+
+
+class RpcBridge:
+    """Minimal helper to run coroutines synchronously inside isolated processes.
+
+    If an event loop is already running, the coroutine is executed on a fresh
+    thread with its own loop to avoid nested run_until_complete errors.
+    """
+
+    def run_sync(self, maybe_coro):
+        if not asyncio.iscoroutine(maybe_coro):
+            return maybe_coro
+
+        try:
+            loop = asyncio.get_running_loop()
+        except RuntimeError:
+            loop = None
+
+        if loop and loop.is_running():
+            result_container = {}
+            exc_container = {}
+
+            def _runner():
+                try:
+                    new_loop = asyncio.new_event_loop()
+                    asyncio.set_event_loop(new_loop)
+                    result_container["value"] = new_loop.run_until_complete(maybe_coro)
+                except Exception as exc:  # pragma: no cover
+                    exc_container["error"] = exc
+                finally:
+                    try:
+                        new_loop.close()
+                    except Exception:
+                        pass
+
+            t = threading.Thread(target=_runner, daemon=True)
+            t.start()
+            t.join()
+
+            if "error" in exc_container:
+                raise exc_container["error"]
+            return result_container.get("value")
+
+        return asyncio.run(maybe_coro)

comfy/isolation/runtime_helpers.py

@@ -0,0 +1,471 @@
+# pylint: disable=consider-using-from-import,import-outside-toplevel,no-member
+from __future__ import annotations
+
+import copy
+import logging
+import os
+from pathlib import Path
+from typing import Any, Dict, List, Set, TYPE_CHECKING
+
+from .proxies.helper_proxies import restore_input_types
+from .shm_forensics import scan_shm_forensics
+
+_IMPORT_TORCH = os.environ.get("PYISOLATE_IMPORT_TORCH", "1") == "1"
+
+_ComfyNodeInternal = object
+latest_io = None
+
+if _IMPORT_TORCH:
+    from comfy_api.internal import _ComfyNodeInternal
+    from comfy_api.latest import _io as latest_io
+
+if TYPE_CHECKING:
+    from .extension_wrapper import ComfyNodeExtension
+
+LOG_PREFIX = "]["
+_PRE_EXEC_MIN_FREE_VRAM_BYTES = 2 * 1024 * 1024 * 1024
+
+
+class _RemoteObjectRegistryCaller:
+    def __init__(self, extension: Any) -> None:
+        self._extension = extension
+
+    def __getattr__(self, method_name: str) -> Any:
+        async def _call(instance_id: str, *args: Any, **kwargs: Any) -> Any:
+            return await self._extension.call_remote_object_method(
+                instance_id,
+                method_name,
+                *args,
+                **kwargs,
+            )
+
+        return _call
+
+
+def _wrap_remote_handles_as_host_proxies(value: Any, extension: Any) -> Any:
+    from pyisolate._internal.remote_handle import RemoteObjectHandle
+
+    if isinstance(value, RemoteObjectHandle):
+        if value.type_name == "ModelPatcher":
+            from comfy.isolation.model_patcher_proxy import ModelPatcherProxy
+
+            proxy = ModelPatcherProxy(value.object_id, manage_lifecycle=False)
+            proxy._rpc_caller = _RemoteObjectRegistryCaller(extension)  # type: ignore[attr-defined]
+            proxy._pyisolate_remote_handle = value  # type: ignore[attr-defined]
+            return proxy
+        if value.type_name == "VAE":
+            from comfy.isolation.vae_proxy import VAEProxy
+
+            proxy = VAEProxy(value.object_id, manage_lifecycle=False)
+            proxy._rpc_caller = _RemoteObjectRegistryCaller(extension)  # type: ignore[attr-defined]
+            proxy._pyisolate_remote_handle = value  # type: ignore[attr-defined]
+            return proxy
+        if value.type_name == "CLIP":
+            from comfy.isolation.clip_proxy import CLIPProxy
+
+            proxy = CLIPProxy(value.object_id, manage_lifecycle=False)
+            proxy._rpc_caller = _RemoteObjectRegistryCaller(extension)  # type: ignore[attr-defined]
+            proxy._pyisolate_remote_handle = value  # type: ignore[attr-defined]
+            return proxy
+        if value.type_name == "ModelSampling":
+            from comfy.isolation.model_sampling_proxy import ModelSamplingProxy
+
+            proxy = ModelSamplingProxy(value.object_id, manage_lifecycle=False)
+            proxy._rpc_caller = _RemoteObjectRegistryCaller(extension)  # type: ignore[attr-defined]
+            proxy._pyisolate_remote_handle = value  # type: ignore[attr-defined]
+            return proxy
+        return value
+
+    if isinstance(value, dict):
+        return {
+            k: _wrap_remote_handles_as_host_proxies(v, extension) for k, v in value.items()
+        }
+
+    if isinstance(value, (list, tuple)):
+        wrapped = [_wrap_remote_handles_as_host_proxies(item, extension) for item in value]
+        return type(value)(wrapped)
+
+    return value
+
+
+def _resource_snapshot() -> Dict[str, int]:
+    fd_count = -1
+    shm_sender_files = 0
+    try:
+        fd_count = len(os.listdir("/proc/self/fd"))
+    except Exception:
+        pass
+    try:
+        shm_root = Path("/dev/shm")
+        if shm_root.exists():
+            prefix = f"torch_{os.getpid()}_"
+            shm_sender_files = sum(1 for _ in shm_root.glob(f"{prefix}*"))
+    except Exception:
+        pass
+    return {"fd_count": fd_count, "shm_sender_files": shm_sender_files}
+
+
+def _tensor_transport_summary(value: Any) -> Dict[str, int]:
+    summary: Dict[str, int] = {
+        "tensor_count": 0,
+        "cpu_tensors": 0,
+        "cuda_tensors": 0,
+        "shared_cpu_tensors": 0,
+        "tensor_bytes": 0,
+    }
+    try:
+        import torch
+    except Exception:
+        return summary
+
+    def visit(node: Any) -> None:
+        if isinstance(node, torch.Tensor):
+            summary["tensor_count"] += 1
+            summary["tensor_bytes"] += int(node.numel() * node.element_size())
+            if node.device.type == "cpu":
+                summary["cpu_tensors"] += 1
+                if node.is_shared():
+                    summary["shared_cpu_tensors"] += 1
+            elif node.device.type == "cuda":
+                summary["cuda_tensors"] += 1
+            return
+        if isinstance(node, dict):
+            for v in node.values():
+                visit(v)
+            return
+        if isinstance(node, (list, tuple)):
+            for v in node:
+                visit(v)
+
+    visit(value)
+    return summary
+
+
+def _extract_hidden_unique_id(inputs: Dict[str, Any]) -> str | None:
+    for key, value in inputs.items():
+        key_text = str(key)
+        if "unique_id" in key_text:
+            return str(value)
+    return None
+
+
+def _flush_tensor_transport_state(marker: str, logger: logging.Logger) -> None:
+    try:
+        from pyisolate import flush_tensor_keeper  # type: ignore[attr-defined]
+    except Exception:
+        return
+    if not callable(flush_tensor_keeper):
+        return
+    flushed = flush_tensor_keeper()
+    if flushed > 0:
+        logger.debug(
+            "%s %s flush_tensor_keeper released=%d", LOG_PREFIX, marker, flushed
+        )
+
+
+def _relieve_host_vram_pressure(marker: str, logger: logging.Logger) -> None:
+    import comfy.model_management as model_management
+
+    model_management.cleanup_models_gc()
+    model_management.cleanup_models()
+
+    device = model_management.get_torch_device()
+    if not hasattr(device, "type") or device.type == "cpu":
+        return
+
+    required = max(
+        model_management.minimum_inference_memory(),
+        _PRE_EXEC_MIN_FREE_VRAM_BYTES,
+    )
+    if model_management.get_free_memory(device) < required:
+        model_management.free_memory(required, device, for_dynamic=True)
+        if model_management.get_free_memory(device) < required:
+            model_management.free_memory(required, device, for_dynamic=False)
+        model_management.cleanup_models()
+        model_management.soft_empty_cache()
+        logger.debug("%s %s free_memory target=%d", LOG_PREFIX, marker, required)
+
+
+def _detach_shared_cpu_tensors(value: Any) -> Any:
+    try:
+        import torch
+    except Exception:
+        return value
+
+    if isinstance(value, torch.Tensor):
+        if value.device.type == "cpu" and value.is_shared():
+            clone = value.clone()
+            if value.requires_grad:
+                clone.requires_grad_(True)
+            return clone
+        return value
+    if isinstance(value, list):
+        return [_detach_shared_cpu_tensors(v) for v in value]
+    if isinstance(value, tuple):
+        return tuple(_detach_shared_cpu_tensors(v) for v in value)
+    if isinstance(value, dict):
+        return {k: _detach_shared_cpu_tensors(v) for k, v in value.items()}
+    return value
+
+
+def build_stub_class(
+    node_name: str,
+    info: Dict[str, object],
+    extension: "ComfyNodeExtension",
+    running_extensions: Dict[str, "ComfyNodeExtension"],
+    logger: logging.Logger,
+) -> type:
+    if latest_io is None:
+        raise RuntimeError("comfy_api.latest._io is required to build isolation stubs")
+    is_v3 = bool(info.get("is_v3", False))
+    function_name = "_pyisolate_execute"
+    restored_input_types = restore_input_types(info.get("input_types", {}))
+
+    async def _execute(self, **inputs):
+        from comfy.isolation import _RUNNING_EXTENSIONS
+
+        # Update BOTH the local dict AND the module-level dict
+        running_extensions[extension.name] = extension
+        _RUNNING_EXTENSIONS[extension.name] = extension
+        prev_child = None
+        node_unique_id = _extract_hidden_unique_id(inputs)
+        summary = _tensor_transport_summary(inputs)
+        resources = _resource_snapshot()
+        logger.debug(
+            "%s ISO:execute_start ext=%s node=%s uid=%s",
+            LOG_PREFIX,
+            extension.name,
+            node_name,
+            node_unique_id or "-",
+        )
+        logger.debug(
+            "%s ISO:execute_start ext=%s node=%s uid=%s tensors=%d cpu=%d cuda=%d shared_cpu=%d bytes=%d fds=%d sender_shm=%d",
+            LOG_PREFIX,
+            extension.name,
+            node_name,
+            node_unique_id or "-",
+            summary["tensor_count"],
+            summary["cpu_tensors"],
+            summary["cuda_tensors"],
+            summary["shared_cpu_tensors"],
+            summary["tensor_bytes"],
+            resources["fd_count"],
+            resources["shm_sender_files"],
+        )
+        scan_shm_forensics("RUNTIME:execute_start", refresh_model_context=True)
+        try:
+            if os.environ.get("PYISOLATE_CHILD") != "1":
+                _relieve_host_vram_pressure("RUNTIME:pre_execute", logger)
+                scan_shm_forensics("RUNTIME:pre_execute", refresh_model_context=True)
+            from pyisolate._internal.model_serialization import (
+                serialize_for_isolation,
+                deserialize_from_isolation,
+            )
+
+            prev_child = os.environ.pop("PYISOLATE_CHILD", None)
+            logger.debug(
+                "%s ISO:serialize_start ext=%s node=%s uid=%s",
+                LOG_PREFIX,
+                extension.name,
+                node_name,
+                node_unique_id or "-",
+            )
+            # Unwrap NodeOutput-like dicts before serialization.
+            # OUTPUT_NODE nodes return {"ui": {...}, "result": (outputs...)}
+            # and the executor may pass this dict as input to downstream nodes.
+            unwrapped_inputs = {}
+            for k, v in inputs.items():
+                if isinstance(v, dict) and "result" in v and ("ui" in v or "__node_output__" in v):
+                    result = v.get("result")
+                    if isinstance(result, (tuple, list)) and len(result) > 0:
+                        unwrapped_inputs[k] = result[0]
+                    else:
+                        unwrapped_inputs[k] = result
+                else:
+                    unwrapped_inputs[k] = v
+            serialized = serialize_for_isolation(unwrapped_inputs)
+            logger.debug(
+                "%s ISO:serialize_done ext=%s node=%s uid=%s",
+                LOG_PREFIX,
+                extension.name,
+                node_name,
+                node_unique_id or "-",
+            )
+            logger.debug(
+                "%s ISO:dispatch_start ext=%s node=%s uid=%s",
+                LOG_PREFIX,
+                extension.name,
+                node_name,
+                node_unique_id or "-",
+            )
+            result = await extension.execute_node(node_name, **serialized)
+            logger.debug(
+                "%s ISO:dispatch_done ext=%s node=%s uid=%s",
+                LOG_PREFIX,
+                extension.name,
+                node_name,
+                node_unique_id or "-",
+            )
+            # Reconstruct NodeOutput if the child serialized one
+            if isinstance(result, dict) and result.get("__node_output__"):
+                from comfy_api.latest import io as latest_io
+                args_raw = result.get("args", ())
+                deserialized_args = await deserialize_from_isolation(args_raw, extension)
+                deserialized_args = _wrap_remote_handles_as_host_proxies(
+                    deserialized_args, extension
+                )
+                deserialized_args = _detach_shared_cpu_tensors(deserialized_args)
+                ui_raw = result.get("ui")
+                deserialized_ui = None
+                if ui_raw is not None:
+                    deserialized_ui = await deserialize_from_isolation(ui_raw, extension)
+                    deserialized_ui = _wrap_remote_handles_as_host_proxies(
+                        deserialized_ui, extension
+                    )
+                    deserialized_ui = _detach_shared_cpu_tensors(deserialized_ui)
+                scan_shm_forensics("RUNTIME:post_execute", refresh_model_context=True)
+                return latest_io.NodeOutput(
+                    *deserialized_args,
+                    ui=deserialized_ui,
+                    expand=result.get("expand"),
+                    block_execution=result.get("block_execution"),
+                )
+            # OUTPUT_NODE: if sealed worker returned a tuple/list whose first
+            # element is a {"ui": ...} dict, unwrap it for the executor.
+            if (isinstance(result, (tuple, list)) and len(result) == 1
+                    and isinstance(result[0], dict) and "ui" in result[0]):
+                return result[0]
+            deserialized = await deserialize_from_isolation(result, extension)
+            deserialized = _wrap_remote_handles_as_host_proxies(deserialized, extension)
+            scan_shm_forensics("RUNTIME:post_execute", refresh_model_context=True)
+            return _detach_shared_cpu_tensors(deserialized)
+        except ImportError:
+            return await extension.execute_node(node_name, **inputs)
+        except Exception:
+            logger.exception(
+                "%s ISO:execute_error ext=%s node=%s uid=%s",
+                LOG_PREFIX,
+                extension.name,
+                node_name,
+                node_unique_id or "-",
+            )
+            raise
+        finally:
+            if prev_child is not None:
+                os.environ["PYISOLATE_CHILD"] = prev_child
+            logger.debug(
+                "%s ISO:execute_end ext=%s node=%s uid=%s",
+                LOG_PREFIX,
+                extension.name,
+                node_name,
+                node_unique_id or "-",
+            )
+            scan_shm_forensics("RUNTIME:execute_end", refresh_model_context=True)
+
+    def _input_types(
+        cls,
+        include_hidden: bool = True,
+        return_schema: bool = False,
+        live_inputs: Any = None,
+    ):
+        if not is_v3:
+            return restored_input_types
+
+        inputs_copy = copy.deepcopy(restored_input_types)
+        if not include_hidden:
+            inputs_copy.pop("hidden", None)
+
+        v3_data: Dict[str, Any] = {"hidden_inputs": {}}
+        dynamic = inputs_copy.pop("dynamic_paths", None)
+        if dynamic is not None:
+            v3_data["dynamic_paths"] = dynamic
+
+        if return_schema:
+            hidden_vals = info.get("hidden", []) or []
+            hidden_enums = []
+            for h in hidden_vals:
+                try:
+                    hidden_enums.append(latest_io.Hidden(h))
+                except Exception:
+                    hidden_enums.append(h)
+
+            class SchemaProxy:
+                hidden = hidden_enums
+
+            return inputs_copy, SchemaProxy, v3_data
+        return inputs_copy
+
+    def _validate_class(cls):
+        return True
+
+    def _get_node_info_v1(cls):
+        node_info = copy.deepcopy(info.get("schema_v1", {}))
+        relative_python_module = node_info.get("python_module")
+        if not isinstance(relative_python_module, str) or not relative_python_module:
+            relative_python_module = f"custom_nodes.{extension.name}"
+        node_info["python_module"] = relative_python_module
+        return node_info
+
+    def _get_base_class(cls):
+        return latest_io.ComfyNode
+
+    attributes: Dict[str, object] = {
+        "FUNCTION": function_name,
+        "CATEGORY": info.get("category", ""),
+        "OUTPUT_NODE": info.get("output_node", False),
+        "RETURN_TYPES": tuple(info.get("return_types", ()) or ()),
+        "RETURN_NAMES": info.get("return_names"),
+        function_name: _execute,
+        "_pyisolate_extension": extension,
+        "_pyisolate_node_name": node_name,
+        "INPUT_TYPES": classmethod(_input_types),
+    }
+
+    output_is_list = info.get("output_is_list")
+    if output_is_list is not None:
+        attributes["OUTPUT_IS_LIST"] = tuple(output_is_list)
+
+    if is_v3:
+        attributes["VALIDATE_CLASS"] = classmethod(_validate_class)
+        attributes["GET_NODE_INFO_V1"] = classmethod(_get_node_info_v1)
+        attributes["GET_BASE_CLASS"] = classmethod(_get_base_class)
+        attributes["DESCRIPTION"] = info.get("description", "")
+        attributes["EXPERIMENTAL"] = info.get("experimental", False)
+        attributes["DEPRECATED"] = info.get("deprecated", False)
+        attributes["API_NODE"] = info.get("api_node", False)
+        attributes["NOT_IDEMPOTENT"] = info.get("not_idempotent", False)
+        attributes["ACCEPT_ALL_INPUTS"] = info.get("accept_all_inputs", False)
+        attributes["_ACCEPT_ALL_INPUTS"] = info.get("accept_all_inputs", False)
+        attributes["INPUT_IS_LIST"] = info.get("input_is_list", False)
+
+    class_name = f"PyIsolate_{node_name}".replace(" ", "_")
+    bases = (_ComfyNodeInternal,) if is_v3 else ()
+    stub_cls = type(class_name, bases, attributes)
+
+    if is_v3:
+        try:
+            stub_cls.VALIDATE_CLASS()
+        except Exception as e:
+            logger.error("%s VALIDATE_CLASS failed: %s - %s", LOG_PREFIX, node_name, e)
+
+    return stub_cls
+
+
+def get_class_types_for_extension(
+    extension_name: str,
+    running_extensions: Dict[str, "ComfyNodeExtension"],
+    specs: List[Any],
+) -> Set[str]:
+    extension = running_extensions.get(extension_name)
+    if not extension:
+        return set()
+
+    ext_path = Path(extension.module_path)
+    class_types = set()
+    for spec in specs:
+        if spec.module_path.resolve() == ext_path.resolve():
+            class_types.add(spec.node_name)
+    return class_types
+
+
+__all__ = ["build_stub_class", "get_class_types_for_extension"]

comfy/isolation/shm_forensics.py

@@ -0,0 +1,217 @@
+# pylint: disable=consider-using-from-import,import-outside-toplevel
+from __future__ import annotations
+
+import atexit
+import hashlib
+import logging
+import os
+from pathlib import Path
+from typing import Any, Dict, List, Set
+
+LOG_PREFIX = "]["
+logger = logging.getLogger(__name__)
+
+
+def _shm_debug_enabled() -> bool:
+    return os.environ.get("COMFY_ISO_SHM_DEBUG") == "1"
+
+
+class _SHMForensicsTracker:
+    def __init__(self) -> None:
+        self._started = False
+        self._tracked_files: Set[str] = set()
+        self._current_model_context: Dict[str, str] = {
+            "id": "unknown",
+            "name": "unknown",
+            "hash": "????",
+        }
+
+    @staticmethod
+    def _snapshot_shm() -> Set[str]:
+        shm_path = Path("/dev/shm")
+        if not shm_path.exists():
+            return set()
+        return {f.name for f in shm_path.glob("torch_*")}
+
+    def start(self) -> None:
+        if self._started or not _shm_debug_enabled():
+            return
+        self._tracked_files = self._snapshot_shm()
+        self._started = True
+        logger.debug(
+            "%s SHM:forensics_enabled tracked=%d", LOG_PREFIX, len(self._tracked_files)
+        )
+
+    def stop(self) -> None:
+        if not self._started:
+            return
+        self.scan("shutdown", refresh_model_context=True)
+        self._started = False
+        logger.debug("%s SHM:forensics_disabled", LOG_PREFIX)
+
+    def _compute_model_hash(self, model_patcher: Any) -> str:
+        try:
+            model_instance_id = getattr(model_patcher, "_instance_id", None)
+            if model_instance_id is not None:
+                model_id_text = str(model_instance_id)
+                return model_id_text[-4:] if len(model_id_text) >= 4 else model_id_text
+
+            import torch
+
+            real_model = (
+                model_patcher.model
+                if hasattr(model_patcher, "model")
+                else model_patcher
+            )
+            tensor = None
+            if hasattr(real_model, "parameters"):
+                for p in real_model.parameters():
+                    if torch.is_tensor(p) and p.numel() > 0:
+                        tensor = p
+                        break
+
+            if tensor is None:
+                return "0000"
+
+            flat = tensor.flatten()
+            values = []
+            indices = [0, flat.shape[0] // 2, flat.shape[0] - 1]
+            for i in indices:
+                if i < flat.shape[0]:
+                    values.append(flat[i].item())
+
+            size = 0
+            if hasattr(model_patcher, "model_size"):
+                size = model_patcher.model_size()
+            sample_str = f"{values}_{id(model_patcher):016x}_{size}"
+            return hashlib.sha256(sample_str.encode()).hexdigest()[-4:]
+        except Exception:
+            return "err!"
+
+    def _get_models_snapshot(self) -> List[Dict[str, Any]]:
+        try:
+            import comfy.model_management as model_management
+        except Exception:
+            return []
+
+        snapshot: List[Dict[str, Any]] = []
+        try:
+            for loaded_model in model_management.current_loaded_models:
+                model = loaded_model.model
+                if model is None:
+                    continue
+                if str(getattr(loaded_model, "device", "")) != "cuda:0":
+                    continue
+
+                name = (
+                    model.model.__class__.__name__
+                    if hasattr(model, "model")
+                    else type(model).__name__
+                )
+                model_hash = self._compute_model_hash(model)
+                model_instance_id = getattr(model, "_instance_id", None)
+                if model_instance_id is None:
+                    model_instance_id = model_hash
+                snapshot.append(
+                    {
+                        "name": str(name),
+                        "id": str(model_instance_id),
+                        "hash": str(model_hash or "????"),
+                        "used": bool(getattr(loaded_model, "currently_used", False)),
+                    }
+                )
+        except Exception:
+            return []
+
+        return snapshot
+
+    def _update_model_context(self) -> None:
+        snapshot = self._get_models_snapshot()
+        selected = None
+
+        used_models = [m for m in snapshot if m.get("used") and m.get("id")]
+        if used_models:
+            selected = used_models[-1]
+        else:
+            live_models = [m for m in snapshot if m.get("id")]
+            if live_models:
+                selected = live_models[-1]
+
+        if selected is None:
+            self._current_model_context = {
+                "id": "unknown",
+                "name": "unknown",
+                "hash": "????",
+            }
+            return
+
+        self._current_model_context = {
+            "id": str(selected.get("id", "unknown")),
+            "name": str(selected.get("name", "unknown")),
+            "hash": str(selected.get("hash", "????") or "????"),
+        }
+
+    def scan(self, marker: str, refresh_model_context: bool = True) -> None:
+        if not self._started or not _shm_debug_enabled():
+            return
+
+        if refresh_model_context:
+            self._update_model_context()
+
+        current = self._snapshot_shm()
+        added = current - self._tracked_files
+        removed = self._tracked_files - current
+        self._tracked_files = current
+
+        if not added and not removed:
+            logger.debug("%s SHM:scan marker=%s changes=0", LOG_PREFIX, marker)
+            return
+
+        for filename in sorted(added):
+            logger.info("%s SHM:created | %s", LOG_PREFIX, filename)
+            model_id = self._current_model_context["id"]
+            if model_id == "unknown":
+                logger.error(
+                    "%s SHM:model_association_missing | file=%s | reason=no_active_model_context",
+                    LOG_PREFIX,
+                    filename,
+                )
+            else:
+                logger.info(
+                    "%s SHM:model_association | model=%s | file=%s | name=%s | hash=%s",
+                    LOG_PREFIX,
+                    model_id,
+                    filename,
+                    self._current_model_context["name"],
+                    self._current_model_context["hash"],
+                )
+
+        for filename in sorted(removed):
+            logger.info("%s SHM:deleted | %s", LOG_PREFIX, filename)
+
+        logger.debug(
+            "%s SHM:scan marker=%s created=%d deleted=%d active=%d",
+            LOG_PREFIX,
+            marker,
+            len(added),
+            len(removed),
+            len(self._tracked_files),
+        )
+
+
+_TRACKER = _SHMForensicsTracker()
+
+
+def start_shm_forensics() -> None:
+    _TRACKER.start()
+
+
+def scan_shm_forensics(marker: str, refresh_model_context: bool = True) -> None:
+    _TRACKER.scan(marker, refresh_model_context=refresh_model_context)
+
+
+def stop_shm_forensics() -> None:
+    _TRACKER.stop()
+
+
+atexit.register(stop_shm_forensics)

comfy/isolation/vae_proxy.py

@@ -0,0 +1,214 @@
+# pylint: disable=attribute-defined-outside-init
+import logging
+from typing import Any
+
+from comfy.isolation.proxies.base import (
+    IS_CHILD_PROCESS,
+    BaseProxy,
+    BaseRegistry,
+    detach_if_grad,
+)
+from comfy.isolation.model_patcher_proxy import ModelPatcherProxy, ModelPatcherRegistry
+
+logger = logging.getLogger(__name__)
+
+
+class FirstStageModelRegistry(BaseRegistry[Any]):
+    _type_prefix = "first_stage_model"
+
+    async def get_property(self, instance_id: str, name: str) -> Any:
+        obj = self._get_instance(instance_id)
+        return getattr(obj, name)
+
+    async def has_property(self, instance_id: str, name: str) -> bool:
+        obj = self._get_instance(instance_id)
+        return hasattr(obj, name)
+
+
+class FirstStageModelProxy(BaseProxy[FirstStageModelRegistry]):
+    _registry_class = FirstStageModelRegistry
+    __module__ = "comfy.ldm.models.autoencoder"
+
+    def __getattr__(self, name: str) -> Any:
+        try:
+            return self._call_rpc("get_property", name)
+        except Exception as e:
+            raise AttributeError(
+                f"'{self.__class__.__name__}' object has no attribute '{name}'"
+            ) from e
+
+    def __repr__(self) -> str:
+        return f"<FirstStageModelProxy {self._instance_id}>"
+
+
+class VAERegistry(BaseRegistry[Any]):
+    _type_prefix = "vae"
+
+    async def get_patcher_id(self, instance_id: str) -> str:
+        vae = self._get_instance(instance_id)
+        return ModelPatcherRegistry().register(vae.patcher)
+
+    async def get_first_stage_model_id(self, instance_id: str) -> str:
+        vae = self._get_instance(instance_id)
+        return FirstStageModelRegistry().register(vae.first_stage_model)
+
+    async def encode(self, instance_id: str, pixels: Any) -> Any:
+        return detach_if_grad(self._get_instance(instance_id).encode(pixels))
+
+    async def encode_tiled(
+        self,
+        instance_id: str,
+        pixels: Any,
+        tile_x: int = 512,
+        tile_y: int = 512,
+        overlap: int = 64,
+    ) -> Any:
+        return detach_if_grad(
+            self._get_instance(instance_id).encode_tiled(
+                pixels, tile_x=tile_x, tile_y=tile_y, overlap=overlap
+            )
+        )
+
+    async def decode(self, instance_id: str, samples: Any, **kwargs: Any) -> Any:
+        return detach_if_grad(self._get_instance(instance_id).decode(samples, **kwargs))
+
+    async def decode_tiled(
+        self,
+        instance_id: str,
+        samples: Any,
+        tile_x: int = 64,
+        tile_y: int = 64,
+        overlap: int = 16,
+        **kwargs: Any,
+    ) -> Any:
+        return detach_if_grad(
+            self._get_instance(instance_id).decode_tiled(
+                samples, tile_x=tile_x, tile_y=tile_y, overlap=overlap, **kwargs
+            )
+        )
+
+    async def get_property(self, instance_id: str, name: str) -> Any:
+        return getattr(self._get_instance(instance_id), name)
+
+    async def memory_used_encode(self, instance_id: str, shape: Any, dtype: Any) -> int:
+        return self._get_instance(instance_id).memory_used_encode(shape, dtype)
+
+    async def memory_used_decode(self, instance_id: str, shape: Any, dtype: Any) -> int:
+        return self._get_instance(instance_id).memory_used_decode(shape, dtype)
+
+    async def process_input(self, instance_id: str, image: Any) -> Any:
+        return detach_if_grad(self._get_instance(instance_id).process_input(image))
+
+    async def process_output(self, instance_id: str, image: Any) -> Any:
+        return detach_if_grad(self._get_instance(instance_id).process_output(image))
+
+
+class VAEProxy(BaseProxy[VAERegistry]):
+    _registry_class = VAERegistry
+    __module__ = "comfy.sd"
+
+    @property
+    def patcher(self) -> ModelPatcherProxy:
+        if not hasattr(self, "_patcher_proxy"):
+            patcher_id = self._call_rpc("get_patcher_id")
+            self._patcher_proxy = ModelPatcherProxy(patcher_id, manage_lifecycle=False)
+        return self._patcher_proxy
+
+    @property
+    def first_stage_model(self) -> FirstStageModelProxy:
+        if not hasattr(self, "_first_stage_model_proxy"):
+            fsm_id = self._call_rpc("get_first_stage_model_id")
+            self._first_stage_model_proxy = FirstStageModelProxy(
+                fsm_id, manage_lifecycle=False
+            )
+        return self._first_stage_model_proxy
+
+    @property
+    def vae_dtype(self) -> Any:
+        return self._get_property("vae_dtype")
+
+    def encode(self, pixels: Any) -> Any:
+        return self._call_rpc("encode", pixels)
+
+    def encode_tiled(
+        self, pixels: Any, tile_x: int = 512, tile_y: int = 512, overlap: int = 64
+    ) -> Any:
+        return self._call_rpc("encode_tiled", pixels, tile_x, tile_y, overlap)
+
+    def decode(self, samples: Any, **kwargs: Any) -> Any:
+        return self._call_rpc("decode", samples, **kwargs)
+
+    def decode_tiled(
+        self,
+        samples: Any,
+        tile_x: int = 64,
+        tile_y: int = 64,
+        overlap: int = 16,
+        **kwargs: Any,
+    ) -> Any:
+        return self._call_rpc(
+            "decode_tiled", samples, tile_x, tile_y, overlap, **kwargs
+        )
+
+    def get_sd(self) -> Any:
+        return self._call_rpc("get_sd")
+
+    def _get_property(self, name: str) -> Any:
+        return self._call_rpc("get_property", name)
+
+    @property
+    def latent_dim(self) -> int:
+        return self._get_property("latent_dim")
+
+    @property
+    def latent_channels(self) -> int:
+        return self._get_property("latent_channels")
+
+    @property
+    def downscale_ratio(self) -> Any:
+        return self._get_property("downscale_ratio")
+
+    @property
+    def upscale_ratio(self) -> Any:
+        return self._get_property("upscale_ratio")
+
+    @property
+    def output_channels(self) -> int:
+        return self._get_property("output_channels")
+
+    @property
+    def not_video(self) -> bool:
+        return self._get_property("not_video")
+
+    @property
+    def device(self) -> Any:
+        return self._get_property("device")
+
+    @property
+    def working_dtypes(self) -> Any:
+        return self._get_property("working_dtypes")
+
+    @property
+    def disable_offload(self) -> bool:
+        return self._get_property("disable_offload")
+
+    @property
+    def size(self) -> Any:
+        return self._get_property("size")
+
+    def memory_used_encode(self, shape: Any, dtype: Any) -> int:
+        return self._call_rpc("memory_used_encode", shape, dtype)
+
+    def memory_used_decode(self, shape: Any, dtype: Any) -> int:
+        return self._call_rpc("memory_used_decode", shape, dtype)
+
+    def process_input(self, image: Any) -> Any:
+        return self._call_rpc("process_input", image)
+
+    def process_output(self, image: Any) -> Any:
+        return self._call_rpc("process_output", image)
+
+
+if not IS_CHILD_PROCESS:
+    _VAE_REGISTRY_SINGLETON = VAERegistry()
+    _FIRST_STAGE_MODEL_REGISTRY_SINGLETON = FirstStageModelRegistry()

comfy/k_diffusion/sampling.py

@@ -1,4 +1,5 @@
 import math
+import os
 from functools import partial
 
 from scipy import integrate
@@ -12,8 +13,8 @@ from . import deis
 from . import sa_solver
 import comfy.model_patcher
 import comfy.model_sampling
-
 import comfy.memory_management
+from comfy.cli_args import args
 from comfy.utils import model_trange as trange
 
 def append_zero(x):
@@ -191,6 +192,13 @@ def sample_euler(model, x, sigmas, extra_args=None, callback=None, disable=None,
     """Implements Algorithm 2 (Euler steps) from Karras et al. (2022)."""
     extra_args = {} if extra_args is None else extra_args
     s_in = x.new_ones([x.shape[0]])
+    isolation_active = args.use_process_isolation or os.environ.get("PYISOLATE_CHILD") == "1"
+    if isolation_active:
+        target_device = sigmas.device
+        if x.device != target_device:
+            x = x.to(target_device)
+        s_in = s_in.to(target_device)
+
     for i in trange(len(sigmas) - 1, disable=disable):
         if s_churn > 0:
             gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.

comfy/latent_formats.py

@@ -224,6 +224,7 @@ class Flux2(LatentFormat):
 
         self.latent_rgb_factors_bias = [-0.0329, -0.0718, -0.0851]
         self.latent_rgb_factors_reshape = lambda t: t.reshape(t.shape[0], 32, 2, 2, t.shape[-2], t.shape[-1]).permute(0, 1, 4, 2, 5, 3).reshape(t.shape[0], 32, t.shape[-2] * 2, t.shape[-1] * 2)
+        self.taesd_decoder_name = "taef2_decoder"
 
     def process_in(self, latent):
         return latent
@@ -783,3 +784,10 @@ class ZImagePixelSpace(ChromaRadiance):
     No VAE encoding/decoding — the model operates directly on RGB pixels.
     """
     pass
+
+class CogVideoX(LatentFormat):
+    latent_channels = 16
+    latent_dimensions = 3
+
+    def __init__(self):
+        self.scale_factor = 1.15258426

comfy/ldm/ace/ace_step15.py

@@ -611,6 +611,7 @@ class AceStepDiTModel(nn.Module):
         intermediate_size,
         patch_size,
         audio_acoustic_hidden_dim,
+        condition_dim=None,
         layer_types=None,
         sliding_window=128,
         rms_norm_eps=1e-6,
@@ -640,7 +641,7 @@ class AceStepDiTModel(nn.Module):
 
         self.time_embed = TimestepEmbedding(256, hidden_size, dtype=dtype, device=device, operations=operations)
         self.time_embed_r = TimestepEmbedding(256, hidden_size, dtype=dtype, device=device, operations=operations)
-        self.condition_embedder = Linear(hidden_size, hidden_size, dtype=dtype, device=device)
+        self.condition_embedder = Linear(condition_dim, hidden_size, dtype=dtype, device=device)
 
         if layer_types is None:
             layer_types = ["full_attention"] * num_layers
@@ -1035,6 +1036,9 @@ class AceStepConditionGenerationModel(nn.Module):
         fsq_dim=2048,
         fsq_levels=[8, 8, 8, 5, 5, 5],
         fsq_input_num_quantizers=1,
+        encoder_hidden_size=2048,
+        encoder_intermediate_size=6144,
+        encoder_num_heads=16,
         audio_model=None,
         dtype=None,
         device=None,
@@ -1054,24 +1058,24 @@ class AceStepConditionGenerationModel(nn.Module):
 
         self.decoder = AceStepDiTModel(
             in_channels, hidden_size, num_dit_layers, num_heads, num_kv_heads, head_dim,
-            intermediate_size, patch_size, audio_acoustic_hidden_dim,
+            intermediate_size, patch_size, audio_acoustic_hidden_dim, condition_dim=encoder_hidden_size,
             layer_types=layer_types, sliding_window=sliding_window, rms_norm_eps=rms_norm_eps,
             dtype=dtype, device=device, operations=operations
         )
         self.encoder = AceStepConditionEncoder(
-            text_hidden_dim, timbre_hidden_dim, hidden_size, num_lyric_layers, num_timbre_layers,
-            num_heads, num_kv_heads, head_dim, intermediate_size, rms_norm_eps,
+            text_hidden_dim, timbre_hidden_dim, encoder_hidden_size, num_lyric_layers, num_timbre_layers,
+            encoder_num_heads, num_kv_heads, head_dim, encoder_intermediate_size, rms_norm_eps,
             dtype=dtype, device=device, operations=operations
         )
         self.tokenizer = AceStepAudioTokenizer(
-            audio_acoustic_hidden_dim, hidden_size, pool_window_size, fsq_dim=fsq_dim, fsq_levels=fsq_levels, fsq_input_num_quantizers=fsq_input_num_quantizers, num_layers=num_tokenizer_layers, head_dim=head_dim, rms_norm_eps=rms_norm_eps,
+            audio_acoustic_hidden_dim, encoder_hidden_size, pool_window_size, fsq_dim=fsq_dim, fsq_levels=fsq_levels, fsq_input_num_quantizers=fsq_input_num_quantizers, num_layers=num_tokenizer_layers, head_dim=head_dim, rms_norm_eps=rms_norm_eps,
             dtype=dtype, device=device, operations=operations
         )
         self.detokenizer = AudioTokenDetokenizer(
-            hidden_size, pool_window_size, audio_acoustic_hidden_dim, num_layers=2, head_dim=head_dim,
+            encoder_hidden_size, pool_window_size, audio_acoustic_hidden_dim, num_layers=2, head_dim=head_dim,
             dtype=dtype, device=device, operations=operations
         )
-        self.null_condition_emb = nn.Parameter(torch.empty(1, 1, hidden_size, dtype=dtype, device=device))
+        self.null_condition_emb = nn.Parameter(torch.empty(1, 1, encoder_hidden_size, dtype=dtype, device=device))
 
     def prepare_condition(
         self,

comfy/ldm/cogvideo/model.py

@@ -0,0 +1,573 @@
+# CogVideoX 3D Transformer - ported to ComfyUI native ops
+# Architecture reference: diffusers CogVideoXTransformer3DModel
+# Style reference: comfy/ldm/wan/model.py
+
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.patcher_extension
+import comfy.ldm.common_dit
+
+
+def _get_1d_rotary_pos_embed(dim, pos, theta=10000.0):
+    """Returns (cos, sin) each with shape [seq_len, dim].
+
+    Frequencies are computed at dim//2 resolution then repeat_interleaved
+    to full dim, matching CogVideoX's interleaved (real, imag) pair format.
+    """
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float32, device=pos.device) / dim))
+    angles = torch.outer(pos.float(), freqs.float())
+    cos = angles.cos().repeat_interleave(2, dim=-1).float()
+    sin = angles.sin().repeat_interleave(2, dim=-1).float()
+    return (cos, sin)
+
+
+def apply_rotary_emb(x, freqs_cos_sin):
+    """Apply CogVideoX rotary embedding to query or key tensor.
+
+    x: [B, heads, seq_len, head_dim]
+    freqs_cos_sin: (cos, sin) each [seq_len, head_dim//2]
+
+    Uses interleaved pair rotation (same as diffusers CogVideoX/Flux).
+    head_dim is reshaped to (-1, 2) pairs, rotated, then flattened back.
+    """
+    cos, sin = freqs_cos_sin
+    cos = cos[None, None, :, :].to(x.device)
+    sin = sin[None, None, :, :].to(x.device)
+
+    # Interleaved pairs: [B, H, S, D] -> [B, H, S, D//2, 2] -> (real, imag)
+    x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)
+    x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+
+    return (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
+
+
+def get_timestep_embedding(timesteps, dim, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1, max_period=10000):
+    half = dim // 2
+    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=timesteps.device) / half)
+    args = timesteps[:, None].float() * freqs[None] * scale
+    embedding = torch.cat([torch.sin(args), torch.cos(args)], dim=-1)
+    if flip_sin_to_cos:
+        embedding = torch.cat([embedding[:, half:], embedding[:, :half]], dim=-1)
+    if dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+def get_3d_sincos_pos_embed(embed_dim, spatial_size, temporal_size, spatial_interpolation_scale=1.0, temporal_interpolation_scale=1.0, device=None):
+    if isinstance(spatial_size, int):
+        spatial_size = (spatial_size, spatial_size)
+
+    grid_w = torch.arange(spatial_size[0], dtype=torch.float32, device=device) / spatial_interpolation_scale
+    grid_h = torch.arange(spatial_size[1], dtype=torch.float32, device=device) / spatial_interpolation_scale
+    grid_t = torch.arange(temporal_size, dtype=torch.float32, device=device) / temporal_interpolation_scale
+
+    grid_t, grid_h, grid_w = torch.meshgrid(grid_t, grid_h, grid_w, indexing="ij")
+
+    embed_dim_spatial = 2 * (embed_dim // 3)
+    embed_dim_temporal = embed_dim // 3
+
+    pos_embed_spatial = _get_2d_sincos_pos_embed(embed_dim_spatial, grid_h, grid_w, device=device)
+    pos_embed_temporal = _get_1d_sincos_pos_embed(embed_dim_temporal, grid_t[:, 0, 0], device=device)
+
+    T, H, W = grid_t.shape
+    pos_embed_temporal = pos_embed_temporal.unsqueeze(1).unsqueeze(1).expand(-1, H, W, -1)
+    pos_embed = torch.cat([pos_embed_temporal, pos_embed_spatial], dim=-1)
+
+    return pos_embed
+
+
+def _get_2d_sincos_pos_embed(embed_dim, grid_h, grid_w, device=None):
+    T, H, W = grid_h.shape
+    half_dim = embed_dim // 2
+    pos_h = _get_1d_sincos_pos_embed(half_dim, grid_h.reshape(-1), device=device).reshape(T, H, W, half_dim)
+    pos_w = _get_1d_sincos_pos_embed(half_dim, grid_w.reshape(-1), device=device).reshape(T, H, W, half_dim)
+    return torch.cat([pos_h, pos_w], dim=-1)
+
+
+def _get_1d_sincos_pos_embed(embed_dim, pos, device=None):
+    half = embed_dim // 2
+    freqs = torch.exp(-math.log(10000.0) * torch.arange(start=0, end=half, dtype=torch.float32, device=device) / half)
+    args = pos.float().reshape(-1)[:, None] * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+    if embed_dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+
+class CogVideoXPatchEmbed(nn.Module):
+    def __init__(self, patch_size=2, patch_size_t=None, in_channels=16, dim=1920,
+                 text_dim=4096, bias=True, sample_width=90, sample_height=60,
+                 sample_frames=49, temporal_compression_ratio=4,
+                 max_text_seq_length=226, spatial_interpolation_scale=1.875,
+                 temporal_interpolation_scale=1.0, use_positional_embeddings=True,
+                 use_learned_positional_embeddings=True,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.patch_size = patch_size
+        self.patch_size_t = patch_size_t
+        self.dim = dim
+        self.sample_height = sample_height
+        self.sample_width = sample_width
+        self.sample_frames = sample_frames
+        self.temporal_compression_ratio = temporal_compression_ratio
+        self.max_text_seq_length = max_text_seq_length
+        self.spatial_interpolation_scale = spatial_interpolation_scale
+        self.temporal_interpolation_scale = temporal_interpolation_scale
+        self.use_positional_embeddings = use_positional_embeddings
+        self.use_learned_positional_embeddings = use_learned_positional_embeddings
+
+        if patch_size_t is None:
+            self.proj = operations.Conv2d(in_channels, dim, kernel_size=patch_size, stride=patch_size, bias=bias, device=device, dtype=dtype)
+        else:
+            self.proj = operations.Linear(in_channels * patch_size * patch_size * patch_size_t, dim, device=device, dtype=dtype)
+
+        self.text_proj = operations.Linear(text_dim, dim, device=device, dtype=dtype)
+
+        if use_positional_embeddings or use_learned_positional_embeddings:
+            persistent = use_learned_positional_embeddings
+            pos_embedding = self._get_positional_embeddings(sample_height, sample_width, sample_frames)
+            self.register_buffer("pos_embedding", pos_embedding, persistent=persistent)
+
+    def _get_positional_embeddings(self, sample_height, sample_width, sample_frames, device=None):
+        post_patch_height = sample_height // self.patch_size
+        post_patch_width = sample_width // self.patch_size
+        post_time_compression_frames = (sample_frames - 1) // self.temporal_compression_ratio + 1
+        if self.patch_size_t is not None:
+            post_time_compression_frames = post_time_compression_frames // self.patch_size_t
+        num_patches = post_patch_height * post_patch_width * post_time_compression_frames
+
+        pos_embedding = get_3d_sincos_pos_embed(
+            self.dim,
+            (post_patch_width, post_patch_height),
+            post_time_compression_frames,
+            self.spatial_interpolation_scale,
+            self.temporal_interpolation_scale,
+            device=device,
+        )
+        pos_embedding = pos_embedding.reshape(-1, self.dim)
+        joint_pos_embedding = pos_embedding.new_zeros(
+            1, self.max_text_seq_length + num_patches, self.dim, requires_grad=False
+        )
+        joint_pos_embedding.data[:, self.max_text_seq_length:].copy_(pos_embedding)
+        return joint_pos_embedding
+
+    def forward(self, text_embeds, image_embeds):
+        input_dtype = text_embeds.dtype
+        text_embeds = self.text_proj(text_embeds.to(self.text_proj.weight.dtype)).to(input_dtype)
+        batch_size, num_frames, channels, height, width = image_embeds.shape
+
+        proj_dtype = self.proj.weight.dtype
+        if self.patch_size_t is None:
+            image_embeds = image_embeds.reshape(-1, channels, height, width)
+            image_embeds = self.proj(image_embeds.to(proj_dtype)).to(input_dtype)
+            image_embeds = image_embeds.view(batch_size, num_frames, *image_embeds.shape[1:])
+            image_embeds = image_embeds.flatten(3).transpose(2, 3)
+            image_embeds = image_embeds.flatten(1, 2)
+        else:
+            p = self.patch_size
+            p_t = self.patch_size_t
+            image_embeds = image_embeds.permute(0, 1, 3, 4, 2)
+            image_embeds = image_embeds.reshape(
+                batch_size, num_frames // p_t, p_t, height // p, p, width // p, p, channels
+            )
+            image_embeds = image_embeds.permute(0, 1, 3, 5, 7, 2, 4, 6).flatten(4, 7).flatten(1, 3)
+            image_embeds = self.proj(image_embeds.to(proj_dtype)).to(input_dtype)
+
+        embeds = torch.cat([text_embeds, image_embeds], dim=1).contiguous()
+
+        if self.use_positional_embeddings or self.use_learned_positional_embeddings:
+            text_seq_length = text_embeds.shape[1]
+            num_image_patches = image_embeds.shape[1]
+
+            if self.use_learned_positional_embeddings:
+                image_pos = self.pos_embedding[
+                    :, self.max_text_seq_length:self.max_text_seq_length + num_image_patches
+                ].to(device=embeds.device, dtype=embeds.dtype)
+            else:
+                image_pos = get_3d_sincos_pos_embed(
+                    self.dim,
+                    (width // self.patch_size, height // self.patch_size),
+                    num_image_patches // ((height // self.patch_size) * (width // self.patch_size)),
+                    self.spatial_interpolation_scale,
+                    self.temporal_interpolation_scale,
+                    device=embeds.device,
+                ).reshape(1, num_image_patches, self.dim).to(dtype=embeds.dtype)
+
+            # Build joint: zeros for text + sincos for image
+            joint_pos = torch.zeros(1, text_seq_length + num_image_patches, self.dim, device=embeds.device, dtype=embeds.dtype)
+            joint_pos[:, text_seq_length:] = image_pos
+            embeds = embeds + joint_pos
+
+        return embeds
+
+
+class CogVideoXLayerNormZero(nn.Module):
+    def __init__(self, time_dim, dim, elementwise_affine=True, eps=1e-5, bias=True,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(time_dim, 6 * dim, bias=bias, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(dim, eps=eps, elementwise_affine=elementwise_affine, device=device, dtype=dtype)
+
+    def forward(self, hidden_states, encoder_hidden_states, temb):
+        shift, scale, gate, enc_shift, enc_scale, enc_gate = self.linear(self.silu(temb)).chunk(6, dim=1)
+        hidden_states = self.norm(hidden_states) * (1 + scale)[:, None, :] + shift[:, None, :]
+        encoder_hidden_states = self.norm(encoder_hidden_states) * (1 + enc_scale)[:, None, :] + enc_shift[:, None, :]
+        return hidden_states, encoder_hidden_states, gate[:, None, :], enc_gate[:, None, :]
+
+
+class CogVideoXAdaLayerNorm(nn.Module):
+    def __init__(self, time_dim, dim, elementwise_affine=True, eps=1e-5,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(time_dim, 2 * dim, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(dim, eps=eps, elementwise_affine=elementwise_affine, device=device, dtype=dtype)
+
+    def forward(self, x, temb):
+        temb = self.linear(self.silu(temb))
+        shift, scale = temb.chunk(2, dim=1)
+        x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]
+        return x
+
+
+class CogVideoXBlock(nn.Module):
+    def __init__(self, dim, num_heads, head_dim, time_dim,
+                 eps=1e-5, ff_inner_dim=None, ff_bias=True,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = head_dim
+
+        self.norm1 = CogVideoXLayerNormZero(time_dim, dim, eps=eps, device=device, dtype=dtype, operations=operations)
+
+        # Self-attention (joint text + latent)
+        self.q = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+        self.k = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+        self.v = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+        self.norm_q = operations.LayerNorm(head_dim, eps=1e-6, elementwise_affine=True, device=device, dtype=dtype)
+        self.norm_k = operations.LayerNorm(head_dim, eps=1e-6, elementwise_affine=True, device=device, dtype=dtype)
+        self.attn_out = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+
+        self.norm2 = CogVideoXLayerNormZero(time_dim, dim, eps=eps, device=device, dtype=dtype, operations=operations)
+
+        # Feed-forward (GELU approximate)
+        inner_dim = ff_inner_dim or dim * 4
+        self.ff_proj = operations.Linear(dim, inner_dim, bias=ff_bias, device=device, dtype=dtype)
+        self.ff_out = operations.Linear(inner_dim, dim, bias=ff_bias, device=device, dtype=dtype)
+
+    def forward(self, hidden_states, encoder_hidden_states, temb, image_rotary_emb=None, transformer_options=None):
+        if transformer_options is None:
+            transformer_options = {}
+        text_seq_length = encoder_hidden_states.size(1)
+
+        # Norm & modulate
+        norm_hidden, norm_encoder, gate_msa, enc_gate_msa = self.norm1(hidden_states, encoder_hidden_states, temb)
+
+        # Joint self-attention
+        qkv_input = torch.cat([norm_encoder, norm_hidden], dim=1)
+        b, s, _ = qkv_input.shape
+        n, d = self.num_heads, self.head_dim
+
+        q = self.q(qkv_input).view(b, s, n, d)
+        k = self.k(qkv_input).view(b, s, n, d)
+        v = self.v(qkv_input)
+
+        q = self.norm_q(q).view(b, s, n, d)
+        k = self.norm_k(k).view(b, s, n, d)
+
+        # Apply rotary embeddings to image tokens only (diffusers format: [B, heads, seq, head_dim])
+        if image_rotary_emb is not None:
+            q_img = q[:, text_seq_length:].transpose(1, 2)  # [B, heads, img_seq, head_dim]
+            k_img = k[:, text_seq_length:].transpose(1, 2)
+            q_img = apply_rotary_emb(q_img, image_rotary_emb)
+            k_img = apply_rotary_emb(k_img, image_rotary_emb)
+            q = torch.cat([q[:, :text_seq_length], q_img.transpose(1, 2)], dim=1)
+            k = torch.cat([k[:, :text_seq_length], k_img.transpose(1, 2)], dim=1)
+
+        attn_out = optimized_attention(
+            q.reshape(b, s, n * d),
+            k.reshape(b, s, n * d),
+            v,
+            heads=self.num_heads,
+            transformer_options=transformer_options,
+        )
+
+        attn_out = self.attn_out(attn_out)
+
+        attn_encoder, attn_hidden = attn_out.split([text_seq_length, s - text_seq_length], dim=1)
+
+        hidden_states = hidden_states + gate_msa * attn_hidden
+        encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_encoder
+
+        # Norm & modulate for FF
+        norm_hidden, norm_encoder, gate_ff, enc_gate_ff = self.norm2(hidden_states, encoder_hidden_states, temb)
+
+        # Feed-forward (GELU on concatenated text + latent)
+        ff_input = torch.cat([norm_encoder, norm_hidden], dim=1)
+        ff_output = self.ff_out(F.gelu(self.ff_proj(ff_input), approximate="tanh"))
+
+        hidden_states = hidden_states + gate_ff * ff_output[:, text_seq_length:]
+        encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_seq_length]
+
+        return hidden_states, encoder_hidden_states
+
+
+class CogVideoXTransformer3DModel(nn.Module):
+    def __init__(self,
+                 num_attention_heads=30,
+                 attention_head_dim=64,
+                 in_channels=16,
+                 out_channels=16,
+                 flip_sin_to_cos=True,
+                 freq_shift=0,
+                 time_embed_dim=512,
+                 ofs_embed_dim=None,
+                 text_embed_dim=4096,
+                 num_layers=30,
+                 dropout=0.0,
+                 attention_bias=True,
+                 sample_width=90,
+                 sample_height=60,
+                 sample_frames=49,
+                 patch_size=2,
+                 patch_size_t=None,
+                 temporal_compression_ratio=4,
+                 max_text_seq_length=226,
+                 spatial_interpolation_scale=1.875,
+                 temporal_interpolation_scale=1.0,
+                 use_rotary_positional_embeddings=False,
+                 use_learned_positional_embeddings=False,
+                 patch_bias=True,
+                 image_model=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+        super().__init__()
+        self.dtype = dtype
+        dim = num_attention_heads * attention_head_dim
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.patch_size = patch_size
+        self.patch_size_t = patch_size_t
+        self.max_text_seq_length = max_text_seq_length
+        self.use_rotary_positional_embeddings = use_rotary_positional_embeddings
+
+        # 1. Patch embedding
+        self.patch_embed = CogVideoXPatchEmbed(
+            patch_size=patch_size,
+            patch_size_t=patch_size_t,
+            in_channels=in_channels,
+            dim=dim,
+            text_dim=text_embed_dim,
+            bias=patch_bias,
+            sample_width=sample_width,
+            sample_height=sample_height,
+            sample_frames=sample_frames,
+            temporal_compression_ratio=temporal_compression_ratio,
+            max_text_seq_length=max_text_seq_length,
+            spatial_interpolation_scale=spatial_interpolation_scale,
+            temporal_interpolation_scale=temporal_interpolation_scale,
+            use_positional_embeddings=not use_rotary_positional_embeddings,
+            use_learned_positional_embeddings=use_learned_positional_embeddings,
+            device=device, dtype=torch.float32, operations=operations,
+        )
+
+        # 2. Time embedding
+        self.time_proj_dim = dim
+        self.time_proj_flip = flip_sin_to_cos
+        self.time_proj_shift = freq_shift
+        self.time_embedding_linear_1 = operations.Linear(dim, time_embed_dim, device=device, dtype=dtype)
+        self.time_embedding_act = nn.SiLU()
+        self.time_embedding_linear_2 = operations.Linear(time_embed_dim, time_embed_dim, device=device, dtype=dtype)
+
+        # Optional OFS embedding (CogVideoX 1.5 I2V)
+        self.ofs_proj_dim = ofs_embed_dim
+        if ofs_embed_dim:
+            self.ofs_embedding_linear_1 = operations.Linear(ofs_embed_dim, ofs_embed_dim, device=device, dtype=dtype)
+            self.ofs_embedding_act = nn.SiLU()
+            self.ofs_embedding_linear_2 = operations.Linear(ofs_embed_dim, ofs_embed_dim, device=device, dtype=dtype)
+        else:
+            self.ofs_embedding_linear_1 = None
+
+        # 3. Transformer blocks
+        self.blocks = nn.ModuleList([
+            CogVideoXBlock(
+                dim=dim,
+                num_heads=num_attention_heads,
+                head_dim=attention_head_dim,
+                time_dim=time_embed_dim,
+                eps=1e-5,
+                device=device, dtype=dtype, operations=operations,
+            )
+            for _ in range(num_layers)
+        ])
+
+        self.norm_final = operations.LayerNorm(dim, eps=1e-5, elementwise_affine=True, device=device, dtype=dtype)
+
+        # 4. Output
+        self.norm_out = CogVideoXAdaLayerNorm(
+            time_dim=time_embed_dim, dim=dim, eps=1e-5,
+            device=device, dtype=dtype, operations=operations,
+        )
+
+        if patch_size_t is None:
+            output_dim = patch_size * patch_size * out_channels
+        else:
+            output_dim = patch_size * patch_size * patch_size_t * out_channels
+
+        self.proj_out = operations.Linear(dim, output_dim, device=device, dtype=dtype)
+
+        self.spatial_interpolation_scale = spatial_interpolation_scale
+        self.temporal_interpolation_scale = temporal_interpolation_scale
+        self.temporal_compression_ratio = temporal_compression_ratio
+
+    def forward(self, x, timestep, context, ofs=None, transformer_options=None, **kwargs):
+        if transformer_options is None:
+            transformer_options = {}
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, ofs, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, ofs=None, transformer_options=None, **kwargs):
+        if transformer_options is None:
+            transformer_options = {}
+        # ComfyUI passes [B, C, T, H, W]
+        batch_size, channels, t, h, w = x.shape
+
+        # Pad to patch size (temporal + spatial), same pattern as WAN
+        p_t = self.patch_size_t if self.patch_size_t is not None else 1
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (p_t, self.patch_size, self.patch_size))
+
+        # CogVideoX expects [B, T, C, H, W]
+        x = x.permute(0, 2, 1, 3, 4)
+        batch_size, num_frames, channels, height, width = x.shape
+
+        # Time embedding
+        t_emb = get_timestep_embedding(timestep, self.time_proj_dim, self.time_proj_flip, self.time_proj_shift)
+        t_emb = t_emb.to(dtype=x.dtype)
+        emb = self.time_embedding_linear_2(self.time_embedding_act(self.time_embedding_linear_1(t_emb)))
+
+        if self.ofs_embedding_linear_1 is not None and ofs is not None:
+            ofs_emb = get_timestep_embedding(ofs, self.ofs_proj_dim, self.time_proj_flip, self.time_proj_shift)
+            ofs_emb = ofs_emb.to(dtype=x.dtype)
+            ofs_emb = self.ofs_embedding_linear_2(self.ofs_embedding_act(self.ofs_embedding_linear_1(ofs_emb)))
+            emb = emb + ofs_emb
+
+        # Patch embedding
+        hidden_states = self.patch_embed(context, x)
+
+        text_seq_length = context.shape[1]
+        encoder_hidden_states = hidden_states[:, :text_seq_length]
+        hidden_states = hidden_states[:, text_seq_length:]
+
+        # Rotary embeddings (if used)
+        image_rotary_emb = None
+        if self.use_rotary_positional_embeddings:
+            post_patch_height = height // self.patch_size
+            post_patch_width = width // self.patch_size
+            if self.patch_size_t is None:
+                post_time = num_frames
+            else:
+                post_time = num_frames // self.patch_size_t
+            image_rotary_emb = self._get_rotary_emb(post_patch_height, post_patch_width, post_time, device=x.device)
+
+        # Transformer blocks
+        for i, block in enumerate(self.blocks):
+            hidden_states, encoder_hidden_states = block(
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                temb=emb,
+                image_rotary_emb=image_rotary_emb,
+                transformer_options=transformer_options,
+            )
+
+        hidden_states = self.norm_final(hidden_states)
+
+        # Output projection
+        hidden_states = self.norm_out(hidden_states, temb=emb)
+        hidden_states = self.proj_out(hidden_states)
+
+        # Unpatchify
+        p = self.patch_size
+        p_t = self.patch_size_t
+
+        if p_t is None:
+            output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, -1, p, p)
+            output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
+        else:
+            output = hidden_states.reshape(
+                batch_size, (num_frames + p_t - 1) // p_t, height // p, width // p, -1, p_t, p, p
+            )
+            output = output.permute(0, 1, 5, 4, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(1, 2)
+
+        # Back to ComfyUI format [B, C, T, H, W] and crop padding
+        output = output.permute(0, 2, 1, 3, 4)[:, :, :t, :h, :w]
+        return output
+
+    def _get_rotary_emb(self, h, w, t, device):
+        """Compute CogVideoX 3D rotary positional embeddings.
+
+        For CogVideoX 1.5 (patch_size_t != None): uses "slice" mode — grid positions
+        are integer arange computed at max_size, then sliced to actual size.
+        For CogVideoX 1.0 (patch_size_t == None): uses "linspace" mode with crop coords
+        scaled by spatial_interpolation_scale.
+        """
+        d = self.attention_head_dim
+        dim_t = d // 4
+        dim_h = d // 8 * 3
+        dim_w = d // 8 * 3
+
+        if self.patch_size_t is not None:
+            # CogVideoX 1.5: "slice" mode — positions are simple integer indices
+            # Compute at max(sample_size, actual_size) then slice to actual
+            base_h = self.patch_embed.sample_height // self.patch_size
+            base_w = self.patch_embed.sample_width // self.patch_size
+            max_h = max(base_h, h)
+            max_w = max(base_w, w)
+
+            grid_h = torch.arange(max_h, device=device, dtype=torch.float32)
+            grid_w = torch.arange(max_w, device=device, dtype=torch.float32)
+            grid_t = torch.arange(t, device=device, dtype=torch.float32)
+        else:
+            # CogVideoX 1.0: "linspace" mode with interpolation scale
+            grid_h = torch.linspace(0, h - 1, h, device=device, dtype=torch.float32) * self.spatial_interpolation_scale
+            grid_w = torch.linspace(0, w - 1, w, device=device, dtype=torch.float32) * self.spatial_interpolation_scale
+            grid_t = torch.arange(t, device=device, dtype=torch.float32)
+
+        freqs_t = _get_1d_rotary_pos_embed(dim_t, grid_t)
+        freqs_h = _get_1d_rotary_pos_embed(dim_h, grid_h)
+        freqs_w = _get_1d_rotary_pos_embed(dim_w, grid_w)
+
+        t_cos, t_sin = freqs_t
+        h_cos, h_sin = freqs_h
+        w_cos, w_sin = freqs_w
+
+        # Slice to actual size (for "slice" mode where grids may be larger)
+        t_cos, t_sin = t_cos[:t], t_sin[:t]
+        h_cos, h_sin = h_cos[:h], h_sin[:h]
+        w_cos, w_sin = w_cos[:w], w_sin[:w]
+
+        # Broadcast and concatenate into [T*H*W, head_dim]
+        t_cos = t_cos[:, None, None, :].expand(-1, h, w, -1)
+        t_sin = t_sin[:, None, None, :].expand(-1, h, w, -1)
+        h_cos = h_cos[None, :, None, :].expand(t, -1, w, -1)
+        h_sin = h_sin[None, :, None, :].expand(t, -1, w, -1)
+        w_cos = w_cos[None, None, :, :].expand(t, h, -1, -1)
+        w_sin = w_sin[None, None, :, :].expand(t, h, -1, -1)
+
+        cos = torch.cat([t_cos, h_cos, w_cos], dim=-1).reshape(t * h * w, -1)
+        sin = torch.cat([t_sin, h_sin, w_sin], dim=-1).reshape(t * h * w, -1)
+        return (cos, sin)

comfy/ldm/cogvideo/vae.py

@@ -0,0 +1,566 @@
+# CogVideoX VAE - ported to ComfyUI native ops
+# Architecture reference: diffusers AutoencoderKLCogVideoX
+# Style reference: comfy/ldm/wan/vae.py
+
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+
+class CausalConv3d(nn.Module):
+    """Causal 3D convolution with temporal padding.
+
+    Uses comfy.ops.Conv3d with autopad='causal_zero' fast path: when input has
+    a single temporal frame and no cache, the 3D conv weight is sliced to act
+    as a 2D conv, avoiding computation on zero-padded temporal dimensions.
+    """
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, pad_mode="constant"):
+        super().__init__()
+        if isinstance(kernel_size, int):
+            kernel_size = (kernel_size,) * 3
+
+        time_kernel, height_kernel, width_kernel = kernel_size
+        self.time_kernel_size = time_kernel
+        self.pad_mode = pad_mode
+
+        height_pad = (height_kernel - 1) // 2
+        width_pad = (width_kernel - 1) // 2
+        self.time_causal_padding = (width_pad, width_pad, height_pad, height_pad, time_kernel - 1, 0)
+
+        stride = stride if isinstance(stride, tuple) else (stride, 1, 1)
+        dilation = (dilation, 1, 1)
+        self.conv = ops.Conv3d(
+            in_channels, out_channels, kernel_size,
+            stride=stride, dilation=dilation,
+            padding=(0, height_pad, width_pad),
+        )
+
+    def forward(self, x, conv_cache=None):
+        if self.pad_mode == "replicate":
+            x = F.pad(x, self.time_causal_padding, mode="replicate")
+            conv_cache = None
+        else:
+            kernel_t = self.time_kernel_size
+            if kernel_t > 1:
+                if conv_cache is None and x.shape[2] == 1:
+                    # Fast path: single frame, no cache. All temporal padding
+                    # frames are copies of the input (replicate-style), so the
+                    # 3D conv reduces to a 2D conv with summed temporal kernel.
+                    w = comfy.ops.cast_to_input(self.conv.weight, x)
+                    b = comfy.ops.cast_to_input(self.conv.bias, x) if self.conv.bias is not None else None
+                    w2d = w.sum(dim=2, keepdim=True)
+                    out = F.conv3d(x, w2d, b,
+                                   self.conv.stride, self.conv.padding,
+                                   self.conv.dilation, self.conv.groups)
+                    return out, None
+                cached = [conv_cache] if conv_cache is not None else [x[:, :, :1]] * (kernel_t - 1)
+                x = torch.cat(cached + [x], dim=2)
+            conv_cache = x[:, :, -self.time_kernel_size + 1:].clone() if self.time_kernel_size > 1 else None
+
+        out = self.conv(x)
+        return out, conv_cache
+
+
+def _interpolate_zq(zq, target_size):
+    """Interpolate latent z to target (T, H, W), matching CogVideoX's first-frame-special handling."""
+    t = target_size[0]
+    if t > 1 and t % 2 == 1:
+        z_first = F.interpolate(zq[:, :, :1], size=(1, target_size[1], target_size[2]))
+        z_rest = F.interpolate(zq[:, :, 1:], size=(t - 1, target_size[1], target_size[2]))
+        return torch.cat([z_first, z_rest], dim=2)
+    return F.interpolate(zq, size=target_size)
+
+
+class SpatialNorm3D(nn.Module):
+    """Spatially conditioned normalization."""
+    def __init__(self, f_channels, zq_channels, groups=32):
+        super().__init__()
+        self.norm_layer = ops.GroupNorm(num_channels=f_channels, num_groups=groups, eps=1e-6, affine=True)
+        self.conv_y = CausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
+        self.conv_b = CausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
+
+    def forward(self, f, zq, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+
+        if zq.shape[-3:] != f.shape[-3:]:
+            zq = _interpolate_zq(zq, f.shape[-3:])
+
+        conv_y, new_cache["conv_y"] = self.conv_y(zq, conv_cache=conv_cache.get("conv_y"))
+        conv_b, new_cache["conv_b"] = self.conv_b(zq, conv_cache=conv_cache.get("conv_b"))
+
+        return self.norm_layer(f) * conv_y + conv_b, new_cache
+
+
+class ResnetBlock3D(nn.Module):
+    """3D ResNet block with optional spatial norm."""
+    def __init__(self, in_channels, out_channels=None, temb_channels=512, groups=32,
+                 eps=1e-6, act_fn="silu", spatial_norm_dim=None, pad_mode="first"):
+        super().__init__()
+        out_channels = out_channels or in_channels
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.spatial_norm_dim = spatial_norm_dim
+
+        if act_fn == "silu":
+            self.nonlinearity = nn.SiLU()
+        elif act_fn == "swish":
+            self.nonlinearity = nn.SiLU()
+        else:
+            self.nonlinearity = nn.SiLU()
+
+        if spatial_norm_dim is None:
+            self.norm1 = ops.GroupNorm(num_channels=in_channels, num_groups=groups, eps=eps)
+            self.norm2 = ops.GroupNorm(num_channels=out_channels, num_groups=groups, eps=eps)
+        else:
+            self.norm1 = SpatialNorm3D(in_channels, spatial_norm_dim, groups=groups)
+            self.norm2 = SpatialNorm3D(out_channels, spatial_norm_dim, groups=groups)
+
+        self.conv1 = CausalConv3d(in_channels, out_channels, kernel_size=3, pad_mode=pad_mode)
+
+        if temb_channels > 0:
+            self.temb_proj = ops.Linear(temb_channels, out_channels)
+
+        self.conv2 = CausalConv3d(out_channels, out_channels, kernel_size=3, pad_mode=pad_mode)
+
+        if in_channels != out_channels:
+            self.conv_shortcut = ops.Conv3d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+        else:
+            self.conv_shortcut = None
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        residual = x
+
+        if zq is not None:
+            x, new_cache["norm1"] = self.norm1(x, zq, conv_cache=conv_cache.get("norm1"))
+        else:
+            x = self.norm1(x)
+
+        x = self.nonlinearity(x)
+        x, new_cache["conv1"] = self.conv1(x, conv_cache=conv_cache.get("conv1"))
+
+        if temb is not None and hasattr(self, "temb_proj"):
+            x = x + self.temb_proj(self.nonlinearity(temb))[:, :, None, None, None]
+
+        if zq is not None:
+            x, new_cache["norm2"] = self.norm2(x, zq, conv_cache=conv_cache.get("norm2"))
+        else:
+            x = self.norm2(x)
+
+        x = self.nonlinearity(x)
+        x, new_cache["conv2"] = self.conv2(x, conv_cache=conv_cache.get("conv2"))
+
+        if self.conv_shortcut is not None:
+            residual = self.conv_shortcut(residual)
+
+        return x + residual, new_cache
+
+
+class Downsample3D(nn.Module):
+    """3D downsampling with optional temporal compression."""
+    def __init__(self, in_channels, out_channels, kernel_size=3, stride=2, padding=0, compress_time=False):
+        super().__init__()
+        self.conv = ops.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, x):
+        if self.compress_time:
+            b, c, t, h, w = x.shape
+            x = x.permute(0, 3, 4, 1, 2).reshape(b * h * w, c, t)
+            if t % 2 == 1:
+                x_first, x_rest = x[..., 0], x[..., 1:]
+                if x_rest.shape[-1] > 0:
+                    x_rest = F.avg_pool1d(x_rest, kernel_size=2, stride=2)
+                x = torch.cat([x_first[..., None], x_rest], dim=-1)
+                x = x.reshape(b, h, w, c, x.shape[-1]).permute(0, 3, 4, 1, 2)
+            else:
+                x = F.avg_pool1d(x, kernel_size=2, stride=2)
+                x = x.reshape(b, h, w, c, x.shape[-1]).permute(0, 3, 4, 1, 2)
+
+        pad = (0, 1, 0, 1)
+        x = F.pad(x, pad, mode="constant", value=0)
+        b, c, t, h, w = x.shape
+        x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+        x = self.conv(x)
+        x = x.reshape(b, t, x.shape[1], x.shape[2], x.shape[3]).permute(0, 2, 1, 3, 4)
+        return x
+
+
+class Upsample3D(nn.Module):
+    """3D upsampling with optional temporal decompression."""
+    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, padding=1, compress_time=False):
+        super().__init__()
+        self.conv = ops.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, x):
+        if self.compress_time:
+            if x.shape[2] > 1 and x.shape[2] % 2 == 1:
+                x_first, x_rest = x[:, :, 0], x[:, :, 1:]
+                x_first = F.interpolate(x_first, scale_factor=2.0)
+                x_rest = F.interpolate(x_rest, scale_factor=2.0)
+                x = torch.cat([x_first[:, :, None, :, :], x_rest], dim=2)
+            elif x.shape[2] > 1:
+                x = F.interpolate(x, scale_factor=2.0)
+            else:
+                x = x.squeeze(2)
+                x = F.interpolate(x, scale_factor=2.0)
+                x = x[:, :, None, :, :]
+        else:
+            b, c, t, h, w = x.shape
+            x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+            x = F.interpolate(x, scale_factor=2.0)
+            x = x.reshape(b, t, c, *x.shape[2:]).permute(0, 2, 1, 3, 4)
+
+        b, c, t, h, w = x.shape
+        x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+        x = self.conv(x)
+        x = x.reshape(b, t, *x.shape[1:]).permute(0, 2, 1, 3, 4)
+        return x
+
+
+class DownBlock3D(nn.Module):
+    def __init__(self, in_channels, out_channels, temb_channels=0, num_layers=1,
+                 eps=1e-6, act_fn="silu", groups=32, add_downsample=True,
+                 compress_time=False, pad_mode="first"):
+        super().__init__()
+        self.resnets = nn.ModuleList([
+            ResnetBlock3D(
+                in_channels=in_channels if i == 0 else out_channels,
+                out_channels=out_channels,
+                temb_channels=temb_channels,
+                groups=groups, eps=eps, act_fn=act_fn, pad_mode=pad_mode,
+            )
+            for i in range(num_layers)
+        ])
+        self.downsamplers = nn.ModuleList([Downsample3D(out_channels, out_channels, compress_time=compress_time)]) if add_downsample else None
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        for i, resnet in enumerate(self.resnets):
+            x, new_cache[f"resnet_{i}"] = resnet(x, temb, zq, conv_cache=conv_cache.get(f"resnet_{i}"))
+        if self.downsamplers is not None:
+            for ds in self.downsamplers:
+                x = ds(x)
+        return x, new_cache
+
+
+class MidBlock3D(nn.Module):
+    def __init__(self, in_channels, temb_channels=0, num_layers=1,
+                 eps=1e-6, act_fn="silu", groups=32, spatial_norm_dim=None, pad_mode="first"):
+        super().__init__()
+        self.resnets = nn.ModuleList([
+            ResnetBlock3D(
+                in_channels=in_channels, out_channels=in_channels,
+                temb_channels=temb_channels, groups=groups, eps=eps,
+                act_fn=act_fn, spatial_norm_dim=spatial_norm_dim, pad_mode=pad_mode,
+            )
+            for _ in range(num_layers)
+        ])
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        for i, resnet in enumerate(self.resnets):
+            x, new_cache[f"resnet_{i}"] = resnet(x, temb, zq, conv_cache=conv_cache.get(f"resnet_{i}"))
+        return x, new_cache
+
+
+class UpBlock3D(nn.Module):
+    def __init__(self, in_channels, out_channels, temb_channels=0, num_layers=1,
+                 eps=1e-6, act_fn="silu", groups=32, spatial_norm_dim=16,
+                 add_upsample=True, compress_time=False, pad_mode="first"):
+        super().__init__()
+        self.resnets = nn.ModuleList([
+            ResnetBlock3D(
+                in_channels=in_channels if i == 0 else out_channels,
+                out_channels=out_channels,
+                temb_channels=temb_channels, groups=groups, eps=eps,
+                act_fn=act_fn, spatial_norm_dim=spatial_norm_dim, pad_mode=pad_mode,
+            )
+            for i in range(num_layers)
+        ])
+        self.upsamplers = nn.ModuleList([Upsample3D(out_channels, out_channels, compress_time=compress_time)]) if add_upsample else None
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        for i, resnet in enumerate(self.resnets):
+            x, new_cache[f"resnet_{i}"] = resnet(x, temb, zq, conv_cache=conv_cache.get(f"resnet_{i}"))
+        if self.upsamplers is not None:
+            for us in self.upsamplers:
+                x = us(x)
+        return x, new_cache
+
+
+class Encoder3D(nn.Module):
+    def __init__(self, in_channels=3, out_channels=16,
+                 block_out_channels=(128, 256, 256, 512),
+                 layers_per_block=3, act_fn="silu",
+                 eps=1e-6, groups=32, pad_mode="first",
+                 temporal_compression_ratio=4):
+        super().__init__()
+        temporal_compress_level = int(np.log2(temporal_compression_ratio))
+
+        self.conv_in = CausalConv3d(in_channels, block_out_channels[0], kernel_size=3, pad_mode=pad_mode)
+
+        self.down_blocks = nn.ModuleList()
+        output_channel = block_out_channels[0]
+        for i in range(len(block_out_channels)):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final = i == len(block_out_channels) - 1
+            compress_time = i < temporal_compress_level
+
+            self.down_blocks.append(DownBlock3D(
+                in_channels=input_channel, out_channels=output_channel,
+                temb_channels=0, num_layers=layers_per_block,
+                eps=eps, act_fn=act_fn, groups=groups,
+                add_downsample=not is_final, compress_time=compress_time,
+            ))
+
+        self.mid_block = MidBlock3D(
+            in_channels=block_out_channels[-1], temb_channels=0,
+            num_layers=2, eps=eps, act_fn=act_fn, groups=groups, pad_mode=pad_mode,
+        )
+
+        self.norm_out = ops.GroupNorm(groups, block_out_channels[-1], eps=1e-6)
+        self.conv_act = nn.SiLU()
+        self.conv_out = CausalConv3d(block_out_channels[-1], 2 * out_channels, kernel_size=3, pad_mode=pad_mode)
+
+    def forward(self, x, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+
+        x, new_cache["conv_in"] = self.conv_in(x, conv_cache=conv_cache.get("conv_in"))
+
+        for i, block in enumerate(self.down_blocks):
+            key = f"down_block_{i}"
+            x, new_cache[key] = block(x, None, None, conv_cache.get(key))
+
+        x, new_cache["mid_block"] = self.mid_block(x, None, None, conv_cache=conv_cache.get("mid_block"))
+
+        x = self.norm_out(x)
+        x = self.conv_act(x)
+        x, new_cache["conv_out"] = self.conv_out(x, conv_cache=conv_cache.get("conv_out"))
+
+        return x, new_cache
+
+
+class Decoder3D(nn.Module):
+    def __init__(self, in_channels=16, out_channels=3,
+                 block_out_channels=(128, 256, 256, 512),
+                 layers_per_block=3, act_fn="silu",
+                 eps=1e-6, groups=32, pad_mode="first",
+                 temporal_compression_ratio=4):
+        super().__init__()
+        reversed_channels = list(reversed(block_out_channels))
+        temporal_compress_level = int(np.log2(temporal_compression_ratio))
+
+        self.conv_in = CausalConv3d(in_channels, reversed_channels[0], kernel_size=3, pad_mode=pad_mode)
+
+        self.mid_block = MidBlock3D(
+            in_channels=reversed_channels[0], temb_channels=0,
+            num_layers=2, eps=eps, act_fn=act_fn, groups=groups,
+            spatial_norm_dim=in_channels, pad_mode=pad_mode,
+        )
+
+        self.up_blocks = nn.ModuleList()
+        output_channel = reversed_channels[0]
+        for i in range(len(block_out_channels)):
+            prev_channel = output_channel
+            output_channel = reversed_channels[i]
+            is_final = i == len(block_out_channels) - 1
+            compress_time = i < temporal_compress_level
+
+            self.up_blocks.append(UpBlock3D(
+                in_channels=prev_channel, out_channels=output_channel,
+                temb_channels=0, num_layers=layers_per_block + 1,
+                eps=eps, act_fn=act_fn, groups=groups,
+                spatial_norm_dim=in_channels,
+                add_upsample=not is_final, compress_time=compress_time,
+            ))
+
+        self.norm_out = SpatialNorm3D(reversed_channels[-1], in_channels, groups=groups)
+        self.conv_act = nn.SiLU()
+        self.conv_out = CausalConv3d(reversed_channels[-1], out_channels, kernel_size=3, pad_mode=pad_mode)
+
+    def forward(self, sample, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+
+        x, new_cache["conv_in"] = self.conv_in(sample, conv_cache=conv_cache.get("conv_in"))
+
+        x, new_cache["mid_block"] = self.mid_block(x, None, sample, conv_cache=conv_cache.get("mid_block"))
+
+        for i, block in enumerate(self.up_blocks):
+            key = f"up_block_{i}"
+            x, new_cache[key] = block(x, None, sample, conv_cache=conv_cache.get(key))
+
+        x, new_cache["norm_out"] = self.norm_out(x, sample, conv_cache=conv_cache.get("norm_out"))
+        x = self.conv_act(x)
+        x, new_cache["conv_out"] = self.conv_out(x, conv_cache=conv_cache.get("conv_out"))
+
+        return x, new_cache
+
+
+
+class AutoencoderKLCogVideoX(nn.Module):
+    """CogVideoX VAE. Spatial tiling/slicing handled by ComfyUI's VAE wrapper.
+
+    Uses rolling temporal decode: conv_in + mid_block + temporal up_blocks run
+    on the full (low-res) tensor, then the expensive spatial-only up_blocks +
+    norm_out + conv_out are processed in small temporal chunks with conv_cache
+    carrying causal state between chunks. This keeps peak VRAM proportional to
+    chunk_size rather than total frame count.
+    """
+
+    def __init__(self,
+                 in_channels=3, out_channels=3,
+                 block_out_channels=(128, 256, 256, 512),
+                 latent_channels=16, layers_per_block=3,
+                 act_fn="silu", eps=1e-6, groups=32,
+                 temporal_compression_ratio=4,
+                 ):
+        super().__init__()
+        self.latent_channels = latent_channels
+        self.temporal_compression_ratio = temporal_compression_ratio
+
+        self.encoder = Encoder3D(
+            in_channels=in_channels, out_channels=latent_channels,
+            block_out_channels=block_out_channels, layers_per_block=layers_per_block,
+            act_fn=act_fn, eps=eps, groups=groups,
+            temporal_compression_ratio=temporal_compression_ratio,
+        )
+        self.decoder = Decoder3D(
+            in_channels=latent_channels, out_channels=out_channels,
+            block_out_channels=block_out_channels, layers_per_block=layers_per_block,
+            act_fn=act_fn, eps=eps, groups=groups,
+            temporal_compression_ratio=temporal_compression_ratio,
+        )
+
+        self.num_latent_frames_batch_size = 2
+        self.num_sample_frames_batch_size = 8
+
+    def encode(self, x):
+        t = x.shape[2]
+        frame_batch = self.num_sample_frames_batch_size
+        remainder = t % frame_batch
+        conv_cache = None
+        enc = []
+
+        # Process remainder frames first so only the first chunk can have an
+        # odd temporal dimension — where Downsample3D's first-frame-special
+        # handling in temporal compression is actually correct.
+        if remainder > 0:
+            chunk, conv_cache = self.encoder(x[:, :, :remainder], conv_cache=conv_cache)
+            enc.append(chunk.to(x.device))
+
+        for start in range(remainder, t, frame_batch):
+            chunk, conv_cache = self.encoder(x[:, :, start:start + frame_batch], conv_cache=conv_cache)
+            enc.append(chunk.to(x.device))
+
+        enc = torch.cat(enc, dim=2)
+        mean, _ = enc.chunk(2, dim=1)
+        return mean
+
+    def decode(self, z):
+        return self._decode_rolling(z)
+
+    def _decode_batched(self, z):
+        """Original batched decode - processes 2 latent frames through full decoder."""
+        t = z.shape[2]
+        frame_batch = self.num_latent_frames_batch_size
+        num_batches = max(t // frame_batch, 1)
+        conv_cache = None
+        dec = []
+        for i in range(num_batches):
+            remaining = t % frame_batch
+            start = frame_batch * i + (0 if i == 0 else remaining)
+            end = frame_batch * (i + 1) + remaining
+            chunk, conv_cache = self.decoder(z[:, :, start:end], conv_cache=conv_cache)
+            dec.append(chunk.cpu())
+        return torch.cat(dec, dim=2).to(z.device)
+
+    def _decode_rolling(self, z):
+        """Rolling decode - processes low-res layers on full tensor, then rolls
+        through expensive high-res layers in temporal chunks."""
+        decoder = self.decoder
+        device = z.device
+
+        # Determine which up_blocks have temporal upsample vs spatial-only.
+        # Temporal up_blocks are cheap (low res), spatial-only are expensive.
+        temporal_compress_level = int(np.log2(self.temporal_compression_ratio))
+        split_at = temporal_compress_level  # first N up_blocks do temporal upsample
+
+        # Phase 1: conv_in + mid_block + temporal up_blocks on full tensor (low/medium res)
+        x, _ = decoder.conv_in(z)
+        x, _ = decoder.mid_block(x, None, z)
+
+        for i in range(split_at):
+            x, _ = decoder.up_blocks[i](x, None, z)
+
+        # Phase 2: remaining spatial-only up_blocks + norm_out + conv_out in temporal chunks
+        remaining_blocks = list(range(split_at, len(decoder.up_blocks)))
+        chunk_size = 4  # pixel frames per chunk through high-res layers
+        t_expanded = x.shape[2]
+
+        if t_expanded <= chunk_size or len(remaining_blocks) == 0:
+            # Small enough to process in one go
+            for i in remaining_blocks:
+                x, _ = decoder.up_blocks[i](x, None, z)
+            x, _ = decoder.norm_out(x, z)
+            x = decoder.conv_act(x)
+            x, _ = decoder.conv_out(x)
+            return x
+
+        # Expand z temporally once to match Phase 2's time dimension.
+        # z stays at latent spatial resolution so this is small (~16 MB vs ~1.3 GB
+        # for the old approach of pre-interpolating to every pixel resolution).
+        z_time_expanded = _interpolate_zq(z, (t_expanded, z.shape[3], z.shape[4]))
+
+        # Process in temporal chunks, interpolating spatially per-chunk to avoid
+        # allocating full [B, C, t_expanded, H, W] tensors at each resolution.
+        dec_out = []
+        conv_caches = {}
+
+        for chunk_start in range(0, t_expanded, chunk_size):
+            chunk_end = min(chunk_start + chunk_size, t_expanded)
+            x_chunk = x[:, :, chunk_start:chunk_end]
+            z_t_chunk = z_time_expanded[:, :, chunk_start:chunk_end]
+            z_spatial_cache = {}
+
+            for i in remaining_blocks:
+                block = decoder.up_blocks[i]
+                cache_key = f"up_block_{i}"
+                hw_key = (x_chunk.shape[3], x_chunk.shape[4])
+                if hw_key not in z_spatial_cache:
+                    if z_t_chunk.shape[3] == hw_key[0] and z_t_chunk.shape[4] == hw_key[1]:
+                        z_spatial_cache[hw_key] = z_t_chunk
+                    else:
+                        z_spatial_cache[hw_key] = F.interpolate(z_t_chunk, size=(z_t_chunk.shape[2], hw_key[0], hw_key[1]))
+                x_chunk, new_cache = block(x_chunk, None, z_spatial_cache[hw_key], conv_cache=conv_caches.get(cache_key))
+                conv_caches[cache_key] = new_cache
+
+            hw_key = (x_chunk.shape[3], x_chunk.shape[4])
+            if hw_key not in z_spatial_cache:
+                z_spatial_cache[hw_key] = F.interpolate(z_t_chunk, size=(z_t_chunk.shape[2], hw_key[0], hw_key[1]))
+            x_chunk, new_cache = decoder.norm_out(x_chunk, z_spatial_cache[hw_key], conv_cache=conv_caches.get("norm_out"))
+            conv_caches["norm_out"] = new_cache
+            x_chunk = decoder.conv_act(x_chunk)
+            x_chunk, new_cache = decoder.conv_out(x_chunk, conv_cache=conv_caches.get("conv_out"))
+            conv_caches["conv_out"] = new_cache
+
+            dec_out.append(x_chunk.cpu())
+            del z_spatial_cache
+
+        del x, z_time_expanded
+        return torch.cat(dec_out, dim=2).to(device)

comfy/ldm/ernie/model.py

@@ -0,0 +1,301 @@
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.model_management
+
+def rope(pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor:
+    assert dim % 2 == 0
+    if not comfy.model_management.supports_fp64(pos.device):
+        device = torch.device("cpu")
+    else:
+        device = pos.device
+
+    scale = torch.arange(0, dim, 2, dtype=torch.float64, device=device) / dim
+    omega = 1.0 / (theta**scale)
+    out = torch.einsum("...n,d->...nd", pos.to(device), omega)
+    out = torch.stack([torch.cos(out), torch.sin(out)], dim=0)
+    return out.to(dtype=torch.float32, device=pos.device)
+
+def apply_rotary_emb(x_in: torch.Tensor, freqs_cis: torch.Tensor) -> torch.Tensor:
+    rot_dim = freqs_cis.shape[-1]
+    x, x_pass = x_in[..., :rot_dim], x_in[..., rot_dim:]
+    cos_ = freqs_cis[0]
+    sin_ = freqs_cis[1]
+    x1, x2 = x.chunk(2, dim=-1)
+    x_rotated = torch.cat((-x2, x1), dim=-1)
+    return torch.cat((x * cos_ + x_rotated * sin_, x_pass), dim=-1)
+
+class ErnieImageEmbedND3(nn.Module):
+    def __init__(self, dim: int, theta: int, axes_dim: tuple):
+        super().__init__()
+        self.dim = dim
+        self.theta = theta
+        self.axes_dim = list(axes_dim)
+
+    def forward(self, ids: torch.Tensor) -> torch.Tensor:
+        emb = torch.cat([rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(3)], dim=-1)
+        emb = emb.unsqueeze(3)  # [2, B, S, 1, head_dim//2]
+        return torch.stack([emb, emb], dim=-1).reshape(*emb.shape[:-1], -1)  # [B, S, 1, head_dim]
+
+class ErnieImagePatchEmbedDynamic(nn.Module):
+    def __init__(self, in_channels: int, embed_dim: int, patch_size: int, operations, device=None, dtype=None):
+        super().__init__()
+        self.patch_size = patch_size
+        self.proj = operations.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size, bias=True, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x)
+        batch_size, dim, height, width = x.shape
+        return x.reshape(batch_size, dim, height * width).transpose(1, 2).contiguous()
+
+class Timesteps(nn.Module):
+    def __init__(self, num_channels: int, flip_sin_to_cos: bool = False):
+        super().__init__()
+        self.num_channels = num_channels
+        self.flip_sin_to_cos = flip_sin_to_cos
+
+    def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
+        half_dim = self.num_channels // 2
+        exponent = -math.log(10000) * torch.arange(half_dim, dtype=torch.float32, device=timesteps.device) / half_dim
+        emb = torch.exp(exponent)
+        emb = timesteps[:, None].float() * emb[None, :]
+        if self.flip_sin_to_cos:
+            emb = torch.cat([torch.cos(emb), torch.sin(emb)], dim=-1)
+        else:
+            emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)
+        return emb
+
+class TimestepEmbedding(nn.Module):
+    def __init__(self, in_channels: int, time_embed_dim: int, operations, device=None, dtype=None):
+        super().__init__()
+        Linear = operations.Linear
+        self.linear_1 = Linear(in_channels, time_embed_dim, bias=True, device=device, dtype=dtype)
+        self.act = nn.SiLU()
+        self.linear_2 = Linear(time_embed_dim, time_embed_dim, bias=True, device=device, dtype=dtype)
+
+    def forward(self, sample: torch.Tensor) -> torch.Tensor:
+        sample = self.linear_1(sample)
+        sample = self.act(sample)
+        sample = self.linear_2(sample)
+        return sample
+
+class ErnieImageAttention(nn.Module):
+    def __init__(self, query_dim: int, heads: int, dim_head: int, eps: float = 1e-6, operations=None, device=None, dtype=None):
+        super().__init__()
+        self.heads = heads
+        self.head_dim = dim_head
+        self.inner_dim = heads * dim_head
+
+        Linear = operations.Linear
+        RMSNorm = operations.RMSNorm
+
+        self.to_q = Linear(query_dim, self.inner_dim, bias=False, device=device, dtype=dtype)
+        self.to_k = Linear(query_dim, self.inner_dim, bias=False, device=device, dtype=dtype)
+        self.to_v = Linear(query_dim, self.inner_dim, bias=False, device=device, dtype=dtype)
+
+        self.norm_q = RMSNorm(dim_head, eps=eps, elementwise_affine=True, device=device, dtype=dtype)
+        self.norm_k = RMSNorm(dim_head, eps=eps, elementwise_affine=True, device=device, dtype=dtype)
+
+        self.to_out = nn.ModuleList([Linear(self.inner_dim, query_dim, bias=False, device=device, dtype=dtype)])
+
+    def forward(self, x: torch.Tensor, attention_mask: torch.Tensor = None, image_rotary_emb: torch.Tensor = None) -> torch.Tensor:
+        B, S, _ = x.shape
+
+        q_flat = self.to_q(x)
+        k_flat = self.to_k(x)
+        v_flat = self.to_v(x)
+
+        query = q_flat.view(B, S, self.heads, self.head_dim)
+        key = k_flat.view(B, S, self.heads, self.head_dim)
+
+        query = self.norm_q(query)
+        key = self.norm_k(key)
+
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb)
+            key = apply_rotary_emb(key, image_rotary_emb)
+
+        q_flat = query.reshape(B, S, -1)
+        k_flat = key.reshape(B, S, -1)
+
+        hidden_states = optimized_attention(q_flat, k_flat, v_flat, self.heads, mask=attention_mask)
+
+        return self.to_out[0](hidden_states)
+
+class ErnieImageFeedForward(nn.Module):
+    def __init__(self, hidden_size: int, ffn_hidden_size: int, operations, device=None, dtype=None):
+        super().__init__()
+        Linear = operations.Linear
+        self.gate_proj = Linear(hidden_size, ffn_hidden_size, bias=False, device=device, dtype=dtype)
+        self.up_proj = Linear(hidden_size, ffn_hidden_size, bias=False, device=device, dtype=dtype)
+        self.linear_fc2 = Linear(ffn_hidden_size, hidden_size, bias=False, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.linear_fc2(self.up_proj(x) * F.gelu(self.gate_proj(x)))
+
+class ErnieImageSharedAdaLNBlock(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, ffn_hidden_size: int, eps: float = 1e-6, operations=None, device=None, dtype=None):
+        super().__init__()
+        RMSNorm = operations.RMSNorm
+
+        self.adaLN_sa_ln = RMSNorm(hidden_size, eps=eps, device=device, dtype=dtype)
+        self.self_attention = ErnieImageAttention(
+            query_dim=hidden_size,
+            dim_head=hidden_size // num_heads,
+            heads=num_heads,
+            eps=eps,
+            operations=operations,
+            device=device,
+            dtype=dtype
+        )
+        self.adaLN_mlp_ln = RMSNorm(hidden_size, eps=eps, device=device, dtype=dtype)
+        self.mlp = ErnieImageFeedForward(hidden_size, ffn_hidden_size, operations=operations, device=device, dtype=dtype)
+
+    def forward(self, x, rotary_pos_emb, temb, attention_mask=None):
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = temb
+
+        residual = x
+        x_norm = self.adaLN_sa_ln(x)
+        x_norm = x_norm * (1 + scale_msa) + shift_msa
+
+        attn_out = self.self_attention(x_norm, attention_mask=attention_mask, image_rotary_emb=rotary_pos_emb)
+        x = residual + gate_msa * attn_out
+
+        residual = x
+        x_norm = self.adaLN_mlp_ln(x)
+        x_norm = x_norm * (1 + scale_mlp) + shift_mlp
+
+        return residual + gate_mlp * self.mlp(x_norm)
+
+class ErnieImageAdaLNContinuous(nn.Module):
+    def __init__(self, hidden_size: int, eps: float = 1e-6, operations=None, device=None, dtype=None):
+        super().__init__()
+        LayerNorm = operations.LayerNorm
+        Linear = operations.Linear
+        self.norm = LayerNorm(hidden_size, elementwise_affine=False, eps=eps, device=device, dtype=dtype)
+        self.linear = Linear(hidden_size, hidden_size * 2, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor, conditioning: torch.Tensor) -> torch.Tensor:
+        scale, shift = self.linear(conditioning).chunk(2, dim=-1)
+        x = self.norm(x)
+        x = torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1))
+        return x
+
+class ErnieImageModel(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int = 4096,
+        num_attention_heads: int = 32,
+        num_layers: int = 36,
+        ffn_hidden_size: int = 12288,
+        in_channels: int = 128,
+        out_channels: int = 128,
+        patch_size: int = 1,
+        text_in_dim: int = 3072,
+        rope_theta: int = 256,
+        rope_axes_dim: tuple = (32, 48, 48),
+        eps: float = 1e-6,
+        qk_layernorm: bool = True,
+        device=None,
+        dtype=None,
+        operations=None,
+        **kwargs
+    ):
+        super().__init__()
+        self.dtype = dtype
+        self.hidden_size = hidden_size
+        self.num_heads = num_attention_heads
+        self.head_dim = hidden_size // num_attention_heads
+        self.patch_size = patch_size
+        self.out_channels = out_channels
+
+        Linear = operations.Linear
+
+        self.x_embedder = ErnieImagePatchEmbedDynamic(in_channels, hidden_size, patch_size, operations, device, dtype)
+        self.text_proj = Linear(text_in_dim, hidden_size, bias=False, device=device, dtype=dtype) if text_in_dim != hidden_size else None
+
+        self.time_proj = Timesteps(hidden_size, flip_sin_to_cos=False)
+        self.time_embedding = TimestepEmbedding(hidden_size, hidden_size, operations, device, dtype)
+
+        self.pos_embed = ErnieImageEmbedND3(dim=self.head_dim, theta=rope_theta, axes_dim=rope_axes_dim)
+
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            Linear(hidden_size, 6 * hidden_size, device=device, dtype=dtype)
+        )
+
+        self.layers = nn.ModuleList([
+            ErnieImageSharedAdaLNBlock(hidden_size, num_attention_heads, ffn_hidden_size, eps, operations, device, dtype)
+            for _ in range(num_layers)
+        ])
+
+        self.final_norm = ErnieImageAdaLNContinuous(hidden_size, eps, operations, device, dtype)
+        self.final_linear = Linear(hidden_size, patch_size * patch_size * out_channels, device=device, dtype=dtype)
+
+    def forward(self, x, timesteps, context, **kwargs):
+        device, dtype = x.device, x.dtype
+        B, C, H, W = x.shape
+        p, Hp, Wp = self.patch_size, H // self.patch_size, W // self.patch_size
+        N_img = Hp * Wp
+
+        img_bsh = self.x_embedder(x)
+
+        text_bth = context
+        if self.text_proj is not None and text_bth.numel() > 0:
+            text_bth = self.text_proj(text_bth)
+        Tmax = text_bth.shape[1]
+
+        hidden_states = torch.cat([img_bsh, text_bth], dim=1)
+
+        text_ids = torch.zeros((B, Tmax, 3), device=device, dtype=torch.float32)
+        text_ids[:, :, 0] = torch.linspace(0, Tmax - 1, steps=Tmax, device=x.device, dtype=torch.float32)
+        index = float(Tmax)
+
+        transformer_options = kwargs.get("transformer_options", {})
+        rope_options = transformer_options.get("rope_options", None)
+
+        h_len, w_len = float(Hp), float(Wp)
+        h_offset, w_offset = 0.0, 0.0
+
+        if rope_options is not None:
+            h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0
+            w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0
+            index += rope_options.get("shift_t", 0.0)
+            h_offset += rope_options.get("shift_y", 0.0)
+            w_offset += rope_options.get("shift_x", 0.0)
+
+        image_ids = torch.zeros((Hp, Wp, 3), device=device, dtype=torch.float32)
+        image_ids[:, :, 0] = image_ids[:, :, 1] + index
+        image_ids[:, :, 1] = image_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=Hp, device=device, dtype=torch.float32).unsqueeze(1)
+        image_ids[:, :, 2] = image_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=Wp, device=device, dtype=torch.float32).unsqueeze(0)
+
+        image_ids = image_ids.view(1, N_img, 3).expand(B, -1, -1)
+
+        rotary_pos_emb = self.pos_embed(torch.cat([image_ids, text_ids], dim=1)).to(x.dtype)
+        del image_ids, text_ids
+
+        sample = self.time_proj(timesteps).to(dtype)
+        c = self.time_embedding(sample)
+
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = [
+            t.unsqueeze(1).contiguous() for t in self.adaLN_modulation(c).chunk(6, dim=-1)
+        ]
+
+        temb = [shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp]
+        for layer in self.layers:
+            hidden_states = layer(hidden_states, rotary_pos_emb, temb)
+
+        hidden_states = self.final_norm(hidden_states, c).type_as(hidden_states)
+
+        patches = self.final_linear(hidden_states)[:, :N_img, :]
+        output = (
+            patches.view(B, Hp, Wp, p, p, self.out_channels)
+            .permute(0, 5, 1, 3, 2, 4)
+            .contiguous()
+            .view(B, self.out_channels, H, W)
+        )
+
+        return output

comfy/ldm/flux/math.py

@@ -16,7 +16,7 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transforme
 
 def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     assert dim % 2 == 0
-    if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled():
+    if not comfy.model_management.supports_fp64(pos.device):
         device = torch.device("cpu")
     else:
         device = pos.device

comfy/ldm/lightricks/vae/audio_vae.py

@@ -4,9 +4,6 @@ import math
 import torch
 import torchaudio
 
-import comfy.model_management
-import comfy.model_patcher
-import comfy.utils as utils
 from comfy.ldm.mmaudio.vae.distributions import DiagonalGaussianDistribution
 from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier
 from comfy.ldm.lightricks.vae.causal_audio_autoencoder import (
@@ -43,30 +40,6 @@ class AudioVAEComponentConfig:
 
         return cls(autoencoder=audio_config, vocoder=vocoder_config)
 
-
-class ModelDeviceManager:
-    """Manages device placement and GPU residency for the composed model."""
-
-    def __init__(self, module: torch.nn.Module):
-        load_device = comfy.model_management.get_torch_device()
-        offload_device = comfy.model_management.vae_offload_device()
-        self.patcher = comfy.model_patcher.ModelPatcher(module, load_device, offload_device)
-
-    def ensure_model_loaded(self) -> None:
-        comfy.model_management.free_memory(
-            self.patcher.model_size(),
-            self.patcher.load_device,
-        )
-        comfy.model_management.load_model_gpu(self.patcher)
-
-    def move_to_load_device(self, tensor: torch.Tensor) -> torch.Tensor:
-        return tensor.to(self.patcher.load_device)
-
-    @property
-    def load_device(self):
-        return self.patcher.load_device
-
-
 class AudioLatentNormalizer:
     """Applies per-channel statistics in patch space and restores original layout."""
 
@@ -132,23 +105,17 @@ class AudioPreprocessor:
 class AudioVAE(torch.nn.Module):
     """High-level Audio VAE wrapper exposing encode and decode entry points."""
 
-    def __init__(self, state_dict: dict, metadata: dict):
+    def __init__(self, metadata: dict):
         super().__init__()
 
         component_config = AudioVAEComponentConfig.from_metadata(metadata)
 
-        vae_sd = utils.state_dict_prefix_replace(state_dict, {"audio_vae.": ""}, filter_keys=True)
-        vocoder_sd = utils.state_dict_prefix_replace(state_dict, {"vocoder.": ""}, filter_keys=True)
-
         self.autoencoder = CausalAudioAutoencoder(config=component_config.autoencoder)
         if "bwe" in component_config.vocoder:
             self.vocoder = VocoderWithBWE(config=component_config.vocoder)
         else:
             self.vocoder = Vocoder(config=component_config.vocoder)
 
-        self.autoencoder.load_state_dict(vae_sd, strict=False)
-        self.vocoder.load_state_dict(vocoder_sd, strict=False)
-
         autoencoder_config = self.autoencoder.get_config()
         self.normalizer = AudioLatentNormalizer(
             AudioPatchifier(
@@ -168,18 +135,12 @@ class AudioVAE(torch.nn.Module):
             n_fft=autoencoder_config["n_fft"],
         )
 
-        self.device_manager = ModelDeviceManager(self)
-
-    def encode(self, audio: dict) -> torch.Tensor:
+    def encode(self, audio, sample_rate=44100) -> torch.Tensor:
         """Encode a waveform dictionary into normalized latent tensors."""
 
-        waveform = audio["waveform"]
-        waveform_sample_rate = audio["sample_rate"]
+        waveform = audio
+        waveform_sample_rate = sample_rate
         input_device = waveform.device
-        # Ensure that Audio VAE is loaded on the correct device.
-        self.device_manager.ensure_model_loaded()
-
-        waveform = self.device_manager.move_to_load_device(waveform)
         expected_channels = self.autoencoder.encoder.in_channels
         if waveform.shape[1] != expected_channels:
             if waveform.shape[1] == 1:
@@ -190,7 +151,7 @@ class AudioVAE(torch.nn.Module):
                 )
 
         mel_spec = self.preprocessor.waveform_to_mel(
-            waveform, waveform_sample_rate, device=self.device_manager.load_device
+            waveform, waveform_sample_rate, device=waveform.device
         )
 
         latents = self.autoencoder.encode(mel_spec)
@@ -204,17 +165,13 @@ class AudioVAE(torch.nn.Module):
         """Decode normalized latent tensors into an audio waveform."""
         original_shape = latents.shape
 
-        # Ensure that Audio VAE is loaded on the correct device.
-        self.device_manager.ensure_model_loaded()
-
-        latents = self.device_manager.move_to_load_device(latents)
         latents = self.normalizer.denormalize(latents)
 
         target_shape = self.target_shape_from_latents(original_shape)
         mel_spec = self.autoencoder.decode(latents, target_shape=target_shape)
 
         waveform = self.run_vocoder(mel_spec)
-        return self.device_manager.move_to_load_device(waveform)
+        return waveform
 
     def target_shape_from_latents(self, latents_shape):
         batch, _, time, _ = latents_shape