ComfyUI v0.15.0

* feat(api-nodes): add ByteDance Seedream-5 model

* made error message more correct

* rename seedream 5.0 model

comfy/ldm/lightricks/embeddings_connector.py

@@ -157,11 +157,9 @@ class Embeddings1DConnector(nn.Module):
         self.num_learnable_registers = num_learnable_registers
         if self.num_learnable_registers:
             self.learnable_registers = nn.Parameter(
-                torch.rand(
+                torch.empty(
                     self.num_learnable_registers, inner_dim, dtype=dtype, device=device
                 )
-                * 2.0
-                - 1.0
             )
 
     def get_fractional_positions(self, indices_grid):

comfy/ops.py

@@ -827,6 +827,10 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 else:
                     sd = {}
 
+                if not hasattr(self, 'weight'):
+                    logging.warning("Warning: state dict on uninitialized op {}".format(prefix))
+                    return sd
+
                 if self.bias is not None:
                     sd["{}bias".format(prefix)] = self.bias
 

comfy/text_encoders/lt.py

@@ -101,6 +101,7 @@ class LTXAVTEModel(torch.nn.Module):
         super().__init__()
         self.dtypes = set()
         self.dtypes.add(dtype)
+        self.compat_mode = False
 
         self.gemma3_12b = Gemma3_12BModel(device=device, dtype=dtype_llama, model_options=model_options, layer="all", layer_idx=None)
         self.dtypes.add(dtype_llama)
@@ -108,6 +109,28 @@ class LTXAVTEModel(torch.nn.Module):
         operations = self.gemma3_12b.operations # TODO
         self.text_embedding_projection = operations.Linear(3840 * 49, 3840, bias=False, dtype=dtype, device=device)
 
+    def enable_compat_mode(self):  # TODO: remove
+        from comfy.ldm.lightricks.embeddings_connector import Embeddings1DConnector
+        operations = self.gemma3_12b.operations
+        dtype = self.text_embedding_projection.weight.dtype
+        device = self.text_embedding_projection.weight.device
+        self.audio_embeddings_connector = Embeddings1DConnector(
+            split_rope=True,
+            double_precision_rope=True,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        self.video_embeddings_connector = Embeddings1DConnector(
+            split_rope=True,
+            double_precision_rope=True,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.compat_mode = True
+
     def set_clip_options(self, options):
         self.execution_device = options.get("execution_device", self.execution_device)
         self.gemma3_12b.set_clip_options(options)
@@ -129,6 +152,12 @@ class LTXAVTEModel(torch.nn.Module):
         out = out.reshape((out.shape[0], out.shape[1], -1))
         out = self.text_embedding_projection(out)
         out = out.float()
+
+        if self.compat_mode:
+            out_vid = self.video_embeddings_connector(out)[0]
+            out_audio = self.audio_embeddings_connector(out)[0]
+            out = torch.concat((out_vid, out_audio), dim=-1)
+
         return out.to(out_device), pooled
 
     def generate(self, tokens, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed):
@@ -152,6 +181,16 @@ class LTXAVTEModel(torch.nn.Module):
                     missing_all.extend([f"{prefix}{k}" for k in missing])
                     unexpected_all.extend([f"{prefix}{k}" for k in unexpected])
 
+            if "model.diffusion_model.audio_embeddings_connector.transformer_1d_blocks.2.attn1.to_q.bias" not in sd:  # TODO: remove
+                ww = sd.get("model.diffusion_model.audio_embeddings_connector.transformer_1d_blocks.0.attn1.to_q.bias", None)
+                if ww is not None:
+                    if ww.shape[0] == 3840:
+                        self.enable_compat_mode()
+                        sdv = comfy.utils.state_dict_prefix_replace(sd, {"model.diffusion_model.video_embeddings_connector.": ""}, filter_keys=True)
+                        self.video_embeddings_connector.load_state_dict(sdv, strict=False, assign=getattr(self, "can_assign_sd", False))
+                        sda = comfy.utils.state_dict_prefix_replace(sd, {"model.diffusion_model.audio_embeddings_connector.": ""}, filter_keys=True)
+                        self.audio_embeddings_connector.load_state_dict(sda, strict=False, assign=getattr(self, "can_assign_sd", False))
+
             return (missing_all, unexpected_all)
 
     def memory_estimation_function(self, token_weight_pairs, device=None):

comfy_api_nodes/apis/bytedance.py

@@ -27,6 +27,7 @@ class Seedream4TaskCreationRequest(BaseModel):
     sequential_image_generation: str = Field("disabled")
     sequential_image_generation_options: Seedream4Options = Field(Seedream4Options(max_images=15))
     watermark: bool = Field(False)
+    output_format: str | None = None
 
 
 class ImageTaskCreationResponse(BaseModel):
@@ -106,6 +107,7 @@ RECOMMENDED_PRESETS_SEEDREAM_4 = [
     ("2496x1664 (3:2)", 2496, 1664),
     ("1664x2496 (2:3)", 1664, 2496),
     ("3024x1296 (21:9)", 3024, 1296),
+    ("3072x3072 (1:1)", 3072, 3072),
     ("4096x4096 (1:1)", 4096, 4096),
     ("Custom", None, None),
 ]

comfy_api_nodes/apis/kling.py

@@ -134,6 +134,13 @@ class ImageToVideoWithAudioRequest(BaseModel):
     shot_type: str | None = Field(None)
 
 
+class KlingAvatarRequest(BaseModel):
+    image: str = Field(...)
+    sound_file: str = Field(...)
+    prompt: str | None = Field(None)
+    mode: str = Field(...)
+
+
 class MotionControlRequest(BaseModel):
     prompt: str = Field(...)
     image_url: str = Field(...)

comfy_api_nodes/nodes_bytedance.py

@@ -37,6 +37,12 @@ from comfy_api_nodes.util import (
 
 BYTEPLUS_IMAGE_ENDPOINT = "/proxy/byteplus/api/v3/images/generations"
 
+SEEDREAM_MODELS = {
+    "seedream 5.0 lite": "seedream-5-0-260128",
+    "seedream-4-5-251128": "seedream-4-5-251128",
+    "seedream-4-0-250828": "seedream-4-0-250828",
+}
+
 # Long-running tasks endpoints(e.g., video)
 BYTEPLUS_TASK_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks"
 BYTEPLUS_TASK_STATUS_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks"  # + /{task_id}
@@ -180,14 +186,13 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="ByteDanceSeedreamNode",
-            display_name="ByteDance Seedream 4.5",
+            display_name="ByteDance Seedream 5.0",
             category="api node/image/ByteDance",
             description="Unified text-to-image generation and precise single-sentence editing at up to 4K resolution.",
             inputs=[
                 IO.Combo.Input(
                     "model",
-                    options=["seedream-4-5-251128", "seedream-4-0-250828"],
-                    tooltip="Model name",
+                    options=list(SEEDREAM_MODELS.keys()),
                 ),
                 IO.String.Input(
                     "prompt",
@@ -198,7 +203,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
                 IO.Image.Input(
                     "image",
                     tooltip="Input image(s) for image-to-image generation. "
-                    "List of 1-10 images for single or multi-reference generation.",
+                    "Reference image(s) for single or multi-reference generation.",
                     optional=True,
                 ),
                 IO.Combo.Input(
@@ -210,8 +215,8 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
                     "width",
                     default=2048,
                     min=1024,
-                    max=4096,
-                    step=8,
+                    max=6240,
+                    step=2,
                     tooltip="Custom width for image. Value is working only if `size_preset` is set to `Custom`",
                     optional=True,
                 ),
@@ -219,8 +224,8 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
                     "height",
                     default=2048,
                     min=1024,
-                    max=4096,
-                    step=8,
+                    max=4992,
+                    step=2,
                     tooltip="Custom height for image. Value is working only if `size_preset` is set to `Custom`",
                     optional=True,
                 ),
@@ -283,7 +288,8 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
                 depends_on=IO.PriceBadgeDepends(widgets=["model"]),
                 expr="""
                 (
-                  $price := $contains(widgets.model, "seedream-4-5-251128") ? 0.04 : 0.03;
+                  $price := $contains(widgets.model, "5.0 lite") ? 0.035 :
+                            $contains(widgets.model, "4-5") ? 0.04 : 0.03;
                   {
                     "type":"usd",
                     "usd": $price,
@@ -309,6 +315,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
         watermark: bool = False,
         fail_on_partial: bool = True,
     ) -> IO.NodeOutput:
+        model = SEEDREAM_MODELS[model]
         validate_string(prompt, strip_whitespace=True, min_length=1)
         w = h = None
         for label, tw, th in RECOMMENDED_PRESETS_SEEDREAM_4:
@@ -318,15 +325,12 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
 
         if w is None or h is None:
             w, h = width, height
-            if not (1024 <= w <= 4096) or not (1024 <= h <= 4096):
-                raise ValueError(
-                    f"Custom size out of range: {w}x{h}. " "Both width and height must be between 1024 and 4096 pixels."
-                )
+
         out_num_pixels = w * h
         mp_provided = out_num_pixels / 1_000_000.0
-        if "seedream-4-5" in model and out_num_pixels < 3686400:
+        if ("seedream-4-5" in model or "seedream-5-0" in model) and out_num_pixels < 3686400:
             raise ValueError(
-                f"Minimum image resolution that Seedream 4.5 can generate is 3.68MP, "
+                f"Minimum image resolution for the selected model is 3.68MP, "
                 f"but {mp_provided:.2f}MP provided."
             )
         if "seedream-4-0" in model and out_num_pixels < 921600:
@@ -334,9 +338,18 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
                 f"Minimum image resolution that the selected model can generate is 0.92MP, "
                 f"but {mp_provided:.2f}MP provided."
             )
+        max_pixels = 10_404_496 if "seedream-5-0" in model else 16_777_216
+        if out_num_pixels > max_pixels:
+            raise ValueError(
+                f"Maximum image resolution for the selected model is {max_pixels / 1_000_000:.2f}MP, "
+                f"but {mp_provided:.2f}MP provided."
+            )
         n_input_images = get_number_of_images(image) if image is not None else 0
-        if n_input_images > 10:
-            raise ValueError(f"Maximum of 10 reference images are supported, but {n_input_images} received.")
+        max_num_of_images = 14 if model == "seedream-5-0-260128" else 10
+        if n_input_images > max_num_of_images:
+            raise ValueError(
+                f"Maximum of {max_num_of_images} reference images are supported, but {n_input_images} received."
+            )
         if sequential_image_generation == "auto" and n_input_images + max_images > 15:
             raise ValueError(
                 "The maximum number of generated images plus the number of reference images cannot exceed 15."
@@ -364,6 +377,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
                 sequential_image_generation=sequential_image_generation,
                 sequential_image_generation_options=Seedream4Options(max_images=max_images),
                 watermark=watermark,
+                output_format="png" if model == "seedream-5-0-260128" else None,
             ),
         )
         if len(response.data) == 1:

comfy_api_nodes/nodes_kling.py

@@ -50,6 +50,7 @@ from comfy_api_nodes.apis import (
 )
 from comfy_api_nodes.apis.kling import (
     ImageToVideoWithAudioRequest,
+    KlingAvatarRequest,
     MotionControlRequest,
     MultiPromptEntry,
     OmniImageParamImage,
@@ -74,6 +75,7 @@ from comfy_api_nodes.util import (
     upload_image_to_comfyapi,
     upload_images_to_comfyapi,
     upload_video_to_comfyapi,
+    validate_audio_duration,
     validate_image_aspect_ratio,
     validate_image_dimensions,
     validate_string,
@@ -3139,6 +3141,103 @@ class KlingFirstLastFrameNode(IO.ComfyNode):
         return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url))
 
 
+class KlingAvatarNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="KlingAvatarNode",
+            display_name="Kling Avatar 2.0",
+            category="api node/video/Kling",
+            description="Generate broadcast-style digital human videos from a single photo and an audio file.",
+            inputs=[
+                IO.Image.Input(
+                    "image",
+                    tooltip="Avatar reference image. "
+                    "Width and height must be at least 300px. Aspect ratio must be between 1:2.5 and 2.5:1.",
+                ),
+                IO.Audio.Input(
+                    "sound_file",
+                    tooltip="Audio input. Must be between 2 and 300 seconds in duration.",
+                ),
+                IO.Combo.Input("mode", options=["std", "pro"]),
+                IO.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    optional=True,
+                    tooltip="Optional prompt to define avatar actions, emotions, and camera movements.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed controls whether the node should re-run; "
+                    "results are non-deterministic regardless of seed.",
+                ),
+            ],
+            outputs=[
+                IO.Video.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["mode"]),
+                expr="""
+                (
+                  $prices := {"std": 0.056, "pro": 0.112};
+                  {"type":"usd","usd": $lookup($prices, widgets.mode), "format":{"suffix":"/second"}}
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        image: Input.Image,
+        sound_file: Input.Audio,
+        mode: str,
+        seed: int,
+        prompt: str = "",
+    ) -> IO.NodeOutput:
+        validate_image_dimensions(image, min_width=300, min_height=300)
+        validate_image_aspect_ratio(image, (1, 2.5), (2.5, 1))
+        validate_audio_duration(sound_file, min_duration=2, max_duration=300)
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/kling/v1/videos/avatar/image2video", method="POST"),
+            response_model=TaskStatusResponse,
+            data=KlingAvatarRequest(
+                image=await upload_image_to_comfyapi(cls, image),
+                sound_file=await upload_audio_to_comfyapi(
+                    cls, sound_file, container_format="mp3", codec_name="libmp3lame", mime_type="audio/mpeg"
+                ),
+                prompt=prompt or None,
+                mode=mode,
+            ),
+        )
+        if response.code:
+            raise RuntimeError(
+                f"Kling request failed. Code: {response.code}, Message: {response.message}, Data: {response.data}"
+            )
+        final_response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/kling/v1/videos/avatar/image2video/{response.data.task_id}"),
+            response_model=TaskStatusResponse,
+            status_extractor=lambda r: (r.data.task_status if r.data else None),
+            max_poll_attempts=800,
+        )
+        return IO.NodeOutput(await download_url_to_video_output(final_response.data.task_result.videos[0].url))
+
+
 class KlingExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
@@ -3167,6 +3266,7 @@ class KlingExtension(ComfyExtension):
             MotionControl,
             KlingVideoNode,
             KlingFirstLastFrameNode,
+            KlingAvatarNode,
         ]
 
 

comfy_extras/nodes_images.py

@@ -6,6 +6,7 @@ import folder_paths
 import json
 import os
 import re
+import math
 import torch
 import comfy.utils
 
@@ -682,6 +683,172 @@ class ImageScaleToMaxDimension(IO.ComfyNode):
     upscale = execute    # TODO: remove
 
 
+class SplitImageToTileList(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="SplitImageToTileList",
+            category="image/batch",
+            search_aliases=["split image", "tile image", "slice image"],
+            display_name="Split Image into List of Tiles",
+            description="Splits an image into a batched list of tiles with a specified overlap.",
+            inputs=[
+                IO.Image.Input("image"),
+                IO.Int.Input("tile_width", default=1024, min=64, max=MAX_RESOLUTION),
+                IO.Int.Input("tile_height", default=1024, min=64, max=MAX_RESOLUTION),
+                IO.Int.Input("overlap", default=128, min=0, max=4096),
+            ],
+            outputs=[
+                IO.Image.Output(is_output_list=True),
+            ],
+        )
+
+    @staticmethod
+    def get_grid_coords(width, height, tile_width, tile_height, overlap):
+        coords = []
+        stride_x = max(1, tile_width - overlap)
+        stride_y = max(1, tile_height - overlap)
+
+        y = 0
+        while y < height:
+            x = 0
+            y_end = min(y + tile_height, height)
+            y_start = max(0, y_end - tile_height)
+
+            while x < width:
+                x_end = min(x + tile_width, width)
+                x_start = max(0, x_end - tile_width)
+
+                coords.append((x_start, y_start, x_end, y_end))
+
+                if x_end >= width:
+                    break
+                x += stride_x
+
+            if y_end >= height:
+                break
+            y += stride_y
+
+        return coords
+
+    @classmethod
+    def execute(cls, image, tile_width, tile_height, overlap):
+        b, h, w, c = image.shape
+        coords = cls.get_grid_coords(w, h, tile_width, tile_height, overlap)
+
+        output_list = []
+        for (x_start, y_start, x_end, y_end) in coords:
+            tile = image[:, y_start:y_end, x_start:x_end, :]
+            output_list.append(tile)
+
+        return IO.NodeOutput(output_list)
+
+
+class ImageMergeTileList(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="ImageMergeTileList",
+            display_name="Merge List of Tiles to Image",
+            category="image/batch",
+            search_aliases=["split image", "tile image", "slice image"],
+            is_input_list=True,
+            inputs=[
+                IO.Image.Input("image_list"),
+                IO.Int.Input("final_width", default=1024, min=64, max=32768),
+                IO.Int.Input("final_height", default=1024, min=64, max=32768),
+                IO.Int.Input("overlap", default=128, min=0, max=4096),
+            ],
+            outputs=[
+                IO.Image.Output(is_output_list=False),
+            ],
+        )
+
+    @staticmethod
+    def get_grid_coords(width, height, tile_width, tile_height, overlap):
+        coords = []
+        stride_x = max(1, tile_width - overlap)
+        stride_y = max(1, tile_height - overlap)
+
+        y = 0
+        while y < height:
+            x = 0
+            y_end = min(y + tile_height, height)
+            y_start = max(0, y_end - tile_height)
+
+            while x < width:
+                x_end = min(x + tile_width, width)
+                x_start = max(0, x_end - tile_width)
+
+                coords.append((x_start, y_start, x_end, y_end))
+
+                if x_end >= width:
+                    break
+                x += stride_x
+
+            if y_end >= height:
+                break
+            y += stride_y
+
+        return coords
+
+    @classmethod
+    def execute(cls, image_list, final_width, final_height, overlap):
+        w = final_width[0]
+        h = final_height[0]
+        ovlp = overlap[0]
+        feather_str = 1.0
+
+        first_tile = image_list[0]
+        b, t_h, t_w, c = first_tile.shape
+        device = first_tile.device
+        dtype = first_tile.dtype
+
+        coords = cls.get_grid_coords(w, h, t_w, t_h, ovlp)
+
+        canvas = torch.zeros((b, h, w, c), device=device, dtype=dtype)
+        weights = torch.zeros((b, h, w, 1), device=device, dtype=dtype)
+
+        if ovlp > 0:
+            y_w = torch.sin(math.pi * torch.linspace(0, 1, t_h, device=device, dtype=dtype))
+            x_w = torch.sin(math.pi * torch.linspace(0, 1, t_w, device=device, dtype=dtype))
+            y_w = torch.clamp(y_w, min=1e-5)
+            x_w = torch.clamp(x_w, min=1e-5)
+
+            sine_mask = (y_w.unsqueeze(1) * x_w.unsqueeze(0)).unsqueeze(0).unsqueeze(-1)
+            flat_mask = torch.ones_like(sine_mask)
+
+            weight_mask = torch.lerp(flat_mask, sine_mask, feather_str)
+        else:
+            weight_mask = torch.ones((1, t_h, t_w, 1), device=device, dtype=dtype)
+
+        for i, (x_start, y_start, x_end, y_end) in enumerate(coords):
+            if i >= len(image_list):
+                break
+
+            tile = image_list[i]
+
+            region_h = y_end - y_start
+            region_w = x_end - x_start
+
+            real_h = min(region_h, tile.shape[1])
+            real_w = min(region_w, tile.shape[2])
+
+            y_end_actual = y_start + real_h
+            x_end_actual = x_start + real_w
+
+            tile_crop = tile[:, :real_h, :real_w, :]
+            mask_crop = weight_mask[:, :real_h, :real_w, :]
+
+            canvas[:, y_start:y_end_actual, x_start:x_end_actual, :] += tile_crop * mask_crop
+            weights[:, y_start:y_end_actual, x_start:x_end_actual, :] += mask_crop
+
+        weights[weights == 0] = 1.0
+        merged_image = canvas / weights
+
+        return IO.NodeOutput(merged_image)
+
+
 class ImagesExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
@@ -701,6 +868,8 @@ class ImagesExtension(ComfyExtension):
             ImageRotate,
             ImageFlip,
             ImageScaleToMaxDimension,
+            SplitImageToTileList,
+            ImageMergeTileList,
         ]
 
 

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.14.1"
+__version__ = "0.15.0"

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.14.1"
+version = "0.15.0"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"

requirements.txt

@@ -1,6 +1,6 @@
-comfyui-frontend-package==1.39.14
-comfyui-workflow-templates==0.8.43
-comfyui-embedded-docs==0.4.1
+comfyui-frontend-package==1.39.16
+comfyui-workflow-templates==0.9.3
+comfyui-embedded-docs==0.4.3
 torch
 torchsde
 torchvision