feat: add model_info output to Load3D node

Signed-off-by: bigcat88 <bigcat88@icloud.com>

comfy/ldm/pixeldit/pid.py

@@ -207,8 +207,9 @@ class PidNet(PixDiT_T2I):
                 f"Flux1/SD3 = 16 channels, Flux2 = 128 channels."
             )
         B = x.shape[0]
-        Hs = x.shape[2] // self.patch_size
-        Ws = x.shape[3] // self.patch_size
+        # Match the backbone's pad_to_patch_size (round up) so the LQ grid lines up with the patch stream.
+        Hs = -(-x.shape[2] // self.patch_size)
+        Ws = -(-x.shape[3] // self.patch_size)
 
         degrade_sigma = degrade_sigma.to(device=x.device, dtype=torch.float32).reshape(-1)
         if degrade_sigma.numel() == 1 and B > 1:

comfy/model_base.py

@@ -1428,6 +1428,23 @@ class PiD(PixelDiTT2I):
             out["degrade_sigma"] = comfy.conds.CONDRegular(degrade_sigma)
         return out
 
+    def resize_cond_for_context_window(self, cond_key, cond_value, window, x_in, device, retain_index_list=[]):
+        if cond_key == "lq_latent" and hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor):
+            lq = cond_value.cond
+            dim = window.dim
+            if dim >= lq.ndim:
+                return None
+            lq_proj = self.diffusion_model.lq_proj
+            ratio = lq_proj.sr_scale * lq_proj.latent_spatial_down_factor
+            # Map x window indices -> lq indices (deduplicated, sorted, in-bounds).
+            lq_size = lq.size(dim)
+            lq_indices = sorted({i // ratio for i in window.index_list if 0 <= i // ratio < lq_size})
+            if not lq_indices:
+                return None
+            idx = tuple([slice(None)] * dim + [lq_indices])
+            return cond_value._copy_with(lq[idx].to(device))
+        return super().resize_cond_for_context_window(cond_key, cond_value, window, x_in, device, retain_index_list=retain_index_list)
+
 
 class WAN21(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):

comfy_api/latest/_io.py

@@ -770,6 +770,23 @@ class Load3DCamera(ComfyTypeIO):
     Type = CameraInfo
 
 
+@comfytype(io_type="LOAD3D_MODEL_INFO")
+class Load3DModelInfo(ComfyTypeIO):
+    class ModelTransform(TypedDict):
+        uuid: str
+        name: str
+        type: str
+        position: dict[str, float | int]
+        rotation: dict[str, float | int | str]
+        quaternion: dict[str, float | int]
+        scale: dict[str, float | int]
+        up: dict[str, float | int]
+        visible: bool
+        matrix: list[float]
+
+    Type = list[ModelTransform]
+
+
 @comfytype(io_type="LOAD_3D")
 class Load3D(ComfyTypeIO):
     """3D models are stored as a dictionary."""
@@ -779,6 +796,7 @@ class Load3D(ComfyTypeIO):
         normal: str
         camera_info: Load3DCamera.CameraInfo
         recording: NotRequired[str]
+        model_info: NotRequired[list[Load3DModelInfo.ModelTransform]]
 
     Type = Model3DDict
 
@@ -2291,6 +2309,7 @@ __all__ = [
     "FlowControl",
     "Accumulation",
     "Load3DCamera",
+    "Load3DModelInfo",
     "Load3D",
     "Load3DAnimation",
     "Photomaker",

comfy_api_nodes/apis/beeble.py

@@ -0,0 +1,32 @@
+from pydantic import BaseModel, Field
+
+
+class CreateSwitchXRequest(BaseModel):
+    generation_type: str = Field(...)
+    source_uri: str = Field(...)
+    alpha_mode: str = Field(...)
+    prompt: str | None = Field(None, max_length=2000)
+    reference_image_uri: str | None = Field(None)
+    alpha_uri: str | None = Field(None)
+    max_resolution: int = Field(1080)
+    callback_url: str | None = Field(None)
+    idempotency_key: str | None = Field(None, max_length=256, min_length=1)
+
+
+class SwitchXOutputUrls(BaseModel):
+    render: str | None = Field(None)
+    source: str | None = Field(None)
+    alpha: str | None = Field(None)
+
+
+class SwitchXStatusResponse(BaseModel):
+    id: str = Field(...)
+    status: str = Field(...)
+    progress: int | None = Field(None)
+    generation_type: str | None = Field(None)
+    alpha_mode: str | None = Field(None)
+    output: SwitchXOutputUrls | None = Field(None)
+    error: str | None = Field(None)
+    created_at: str | None = Field(None)
+    modified_at: str | None = Field(None)
+    completed_at: str | None = Field(None)

comfy_api_nodes/nodes_beeble.py

@@ -0,0 +1,404 @@
+from fractions import Fraction
+
+from typing_extensions import override
+
+from comfy_api.latest import IO, ComfyExtension, Input, InputImpl, Types
+from comfy_api_nodes.apis.beeble import (
+    CreateSwitchXRequest,
+    SwitchXStatusResponse,
+)
+from comfy_api_nodes.util import (
+    ApiEndpoint,
+    bytesio_to_image_tensor,
+    convert_mask_to_image,
+    download_url_as_bytesio,
+    download_url_to_image_tensor,
+    download_url_to_video_output,
+    downscale_image_tensor,
+    downscale_video_to_max_pixels,
+    poll_op,
+    sync_op,
+    upload_image_to_comfyapi,
+    upload_video_to_comfyapi,
+    validate_string,
+    validate_video_frame_count,
+)
+
+_MAX_PIXELS = 2_770_000
+_MAX_FRAMES = 240
+_MAX_PROMPT_LEN = 2000
+
+
+def _validate_inputs(prompt: str | None, reference_image: Input.Image | None) -> str | None:
+    """Beeble requires at least one of prompt or reference_image. Returns the cleaned prompt."""
+    cleaned = prompt.strip() if prompt else ""
+    if not cleaned and reference_image is None:
+        raise ValueError("At least one of 'prompt' or 'reference_image' must be provided.")
+    if cleaned:
+        validate_string(cleaned, strip_whitespace=False, max_length=_MAX_PROMPT_LEN)
+    return cleaned or None
+
+
+async def _upload_mask_as_image(
+    cls: type[IO.ComfyNode],
+    mask: Input.Image,
+    *,
+    wait_label: str,
+) -> str:
+    """Encode a single-frame MASK (H, W) or (1, H, W) as a PNG and upload."""
+    if mask.dim() == 2:
+        mask = mask.unsqueeze(0)
+    image = convert_mask_to_image(mask[:1])
+    return await upload_image_to_comfyapi(
+        cls,
+        image,
+        mime_type="image/png",
+        wait_label=wait_label,
+        total_pixels=_MAX_PIXELS,
+    )
+
+
+async def _upload_mask_batch_as_video(
+    cls: type[IO.ComfyNode],
+    mask: Input.Image,
+    *,
+    frame_rate: Fraction,
+    source_frame_count: int,
+    wait_label: str,
+) -> str:
+    """Encode a MASK batch (N, H, W) as a grayscale H.264 MP4 at frame_rate and upload.
+
+    The matte is always downscaled to the pixel budget so it stays within Beeble's limit and
+    keeps the same dimensions as the (similarly downscaled) source — both use the same algorithm
+    from the same starting dimensions, and downscaling is a no-op when already within budget.
+    """
+    if mask.dim() == 2:
+        mask = mask.unsqueeze(0)
+    if mask.shape[0] != source_frame_count:
+        raise ValueError(
+            f"Custom alpha video frame count ({mask.shape[0]}) does not match the "
+            f"source video frame count ({source_frame_count}). The Beeble API requires "
+            "one mask per source frame."
+        )
+    images = downscale_image_tensor(convert_mask_to_image(mask), _MAX_PIXELS)
+    alpha_video = InputImpl.VideoFromComponents(Types.VideoComponents(images=images, audio=None, frame_rate=frame_rate))
+    return await upload_video_to_comfyapi(cls, alpha_video, wait_label=wait_label)
+
+
+def _alpha_mode_input(*, video: bool) -> IO.DynamicCombo.Input:
+    """Build the alpha_mode DynamicCombo with mode-specific extra inputs."""
+    select_keyframe_tooltip = (
+        "First-frame keyframe mask. Beeble propagates this across the video." if video else "Grayscale keyframe mask."
+    )
+    custom_tooltip = (
+        "Per-frame grayscale mask covering the entire video. "
+        "Must have the same frame count as the source. "
+        "Connect a MASK output from SAM3_TrackToMask or similar."
+        if video
+        else "Grayscale mask to apply."
+    )
+    return IO.DynamicCombo.Input(
+        "alpha_mode",
+        tooltip=(
+            "Controls how SwitchX decides what to keep vs. regenerate. "
+            "'auto' isolates the main subject automatically. "
+            "'fill' regenerates the entire frame while preserving geometry. "
+            "'select' propagates a first-frame keyframe across the clip. "
+            "'custom' uses a per-frame alpha matte you provide."
+        ),
+        options=[
+            IO.DynamicCombo.Option("auto", []),
+            IO.DynamicCombo.Option("fill", []),
+            IO.DynamicCombo.Option(
+                "select",
+                [IO.Mask.Input("alpha_keyframe", tooltip=select_keyframe_tooltip)],
+            ),
+            IO.DynamicCombo.Option(
+                "custom",
+                [IO.Mask.Input("alpha_mask", tooltip=custom_tooltip)],
+            ),
+        ],
+    )
+
+
+def _common_inputs(*, source: IO.Input, video: bool) -> list[IO.Input]:
+    return [
+        source,
+        IO.String.Input(
+            "prompt",
+            multiline=True,
+            default="",
+            tooltip=(
+                "Text description of the desired output (max 2000 chars). "
+                "At least one of 'prompt' or 'reference_image' is required."
+            ),
+        ),
+        IO.Image.Input(
+            "reference_image",
+            optional=True,
+            tooltip=(
+                "Reference image whose look (background, lighting, costume) the result "
+                "should adopt. At least one of 'reference_image' or 'prompt' is required."
+            ),
+        ),
+        _alpha_mode_input(video=video),
+        IO.Combo.Input(
+            "max_resolution",
+            options=["1080p", "720p"],
+            default="1080p",
+            tooltip="Maximum output resolution.",
+        ),
+        IO.Int.Input(
+            "seed",
+            default=0,
+            min=0,
+            max=2147483647,
+            control_after_generate=True,
+            tooltip=(
+                "Seed controls whether the node should re-run; " "results are non-deterministic regardless of seed."
+            ),
+        ),
+    ]
+
+
+async def _submit_and_poll(
+    cls: type[IO.ComfyNode],
+    request: CreateSwitchXRequest,
+) -> SwitchXStatusResponse:
+    initial = await sync_op(
+        cls,
+        ApiEndpoint(path="/proxy/beeble/v1/switchx/generations", method="POST"),
+        response_model=SwitchXStatusResponse,
+        data=request,
+    )
+    return await poll_op(
+        cls,
+        ApiEndpoint(path=f"/proxy/beeble/v1/switchx/generations/{initial.id}"),
+        response_model=SwitchXStatusResponse,
+        status_extractor=lambda r: r.status,
+        progress_extractor=lambda r: r.progress,
+    )
+
+
+def _require_output_url(response: SwitchXStatusResponse, name: str) -> str:
+    if response.output is None or getattr(response.output, name) is None:
+        raise RuntimeError(f"Beeble job {response.id} completed without a {name!r} output URL.")
+    return getattr(response.output, name)
+
+
+def _alpha_url(response: SwitchXStatusResponse, mode: str) -> str | None:
+    """URL of the alpha matte, or None when the mode produces no separate matte.
+
+    'fill' selects the whole frame, so Beeble writes no alpha asset even though the status
+    response still returns a (dangling) signed URL for it — fetching it 403s with S3
+    AccessDenied. The other three modes ('auto', 'custom', 'select') all produce a real,
+    downloadable matte.
+    """
+    if mode == "fill" or response.output is None:
+        return None
+    return response.output.alpha
+
+
+class BeebleSwitchXVideoEdit(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="BeebleSwitchXVideoEdit",
+            display_name="Beeble SwitchX Video Edit",
+            category="api node/video/Beeble",
+            description=(
+                "Edit a video with Beeble SwitchX. Switches anything in the scene (background, "
+                "lighting, costume) while preserving the original subject's pixels and motion. "
+                "Provide a reference image and/or text prompt to describe the new look. "
+                "Max 240 frames, max ~2.77MP per frame."
+            ),
+            inputs=_common_inputs(source=IO.Video.Input("video"), video=True),
+            outputs=[
+                IO.Video.Output(display_name="video"),
+                IO.Video.Output(
+                    display_name="alpha",
+                    tooltip="The alpha matte Beeble used. Empty for 'fill' mode, which has no separate matte.",
+                ),
+            ],
+            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=["max_resolution"]),
+                expr="""
+                (
+                  $rate := widgets.max_resolution = "1080p" ? 0.429 : 0.143;
+                  {"type":"usd","usd": $rate, "format":{"suffix":"/30 frames"}}
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        video: Input.Video,
+        prompt: str,
+        alpha_mode: dict,
+        max_resolution: str,
+        seed: int,
+        reference_image: Input.Image | None = None,
+    ) -> IO.NodeOutput:
+        cleaned_prompt = _validate_inputs(prompt, reference_image)
+
+        validate_video_frame_count(video, max_frame_count=_MAX_FRAMES)
+        video = downscale_video_to_max_pixels(video, _MAX_PIXELS)
+
+        mode = alpha_mode["alpha_mode"]
+        alpha_uri: str | None = None
+        if mode == "select":
+            alpha_uri = await _upload_mask_as_image(cls, alpha_mode["alpha_keyframe"], wait_label="Uploading keyframe")
+        elif mode == "custom":
+            alpha_uri = await _upload_mask_batch_as_video(
+                cls,
+                alpha_mode["alpha_mask"],
+                frame_rate=video.get_frame_rate(),
+                source_frame_count=video.get_frame_count(),
+                wait_label="Uploading alpha video",
+            )
+
+        source_uri = await upload_video_to_comfyapi(cls, video, wait_label="Uploading source")
+        reference_uri: str | None = None
+        if reference_image is not None:
+            reference_uri = await upload_image_to_comfyapi(
+                cls,
+                reference_image,
+                mime_type="image/png",
+                wait_label="Uploading reference",
+                total_pixels=_MAX_PIXELS,
+            )
+
+        request = CreateSwitchXRequest(
+            generation_type="video",
+            source_uri=source_uri,
+            alpha_mode=mode,
+            prompt=cleaned_prompt,
+            reference_image_uri=reference_uri,
+            alpha_uri=alpha_uri,
+            max_resolution=1080 if max_resolution == "1080p" else 720,
+        )
+        response = await _submit_and_poll(cls, request)
+
+        render = await download_url_to_video_output(_require_output_url(response, "render"))
+        alpha = None
+        if (alpha_url := _alpha_url(response, mode)) is not None:
+            alpha = await download_url_to_video_output(alpha_url)
+        return IO.NodeOutput(render, alpha)
+
+
+class BeebleSwitchXImageEdit(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="BeebleSwitchXImageEdit",
+            display_name="Beeble SwitchX Image Edit",
+            category="api node/image/Beeble",
+            description=(
+                "Edit a single image with Beeble SwitchX. Switches anything in the scene "
+                "(background, lighting, costume) while preserving the original subject's pixels. "
+                "Provide a reference image and/or text prompt to describe the new look. "
+                "Max ~2.77MP."
+            ),
+            inputs=_common_inputs(source=IO.Image.Input("image"), video=False),
+            outputs=[
+                IO.Image.Output(display_name="image"),
+                IO.Mask.Output(
+                    display_name="alpha",
+                    tooltip="The alpha matte Beeble used. Empty for 'fill' mode, which has no separate matte.",
+                ),
+            ],
+            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=["max_resolution"]),
+                expr="""
+                (
+                  $rate := widgets.max_resolution = "1080p" ? 0.429 : 0.143;
+                  {"type":"usd","usd": $rate}
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        image: Input.Image,
+        prompt: str,
+        alpha_mode: dict,
+        max_resolution: str,
+        seed: int,
+        reference_image: Input.Image | None = None,
+    ) -> IO.NodeOutput:
+        cleaned_prompt = _validate_inputs(prompt, reference_image)
+
+        image = downscale_image_tensor(image, _MAX_PIXELS)
+
+        mode = alpha_mode["alpha_mode"]
+        alpha_uri: str | None = None
+        if mode == "select":
+            alpha_uri = await _upload_mask_as_image(cls, alpha_mode["alpha_keyframe"], wait_label="Uploading keyframe")
+        elif mode == "custom":
+            alpha_uri = await _upload_mask_as_image(cls, alpha_mode["alpha_mask"], wait_label="Uploading alpha")
+
+        source_uri = await upload_image_to_comfyapi(
+            cls,
+            image,
+            mime_type="image/png",
+            wait_label="Uploading source",
+            total_pixels=None,
+        )
+        reference_uri: str | None = None
+        if reference_image is not None:
+            reference_uri = await upload_image_to_comfyapi(
+                cls,
+                reference_image,
+                mime_type="image/png",
+                wait_label="Uploading reference",
+                total_pixels=_MAX_PIXELS,
+            )
+
+        request = CreateSwitchXRequest(
+            generation_type="image",
+            source_uri=source_uri,
+            alpha_mode=mode,
+            prompt=cleaned_prompt,
+            reference_image_uri=reference_uri,
+            alpha_uri=alpha_uri,
+            max_resolution=1080 if max_resolution == "1080p" else 720,
+        )
+        response = await _submit_and_poll(cls, request)
+
+        render = await download_url_to_image_tensor(_require_output_url(response, "render"))
+        alpha_mask = None
+        if (alpha_url := _alpha_url(response, mode)) is not None:
+            alpha_image = bytesio_to_image_tensor(await download_url_as_bytesio(alpha_url), mode="L")
+            alpha_mask = alpha_image.squeeze(-1) if alpha_image.dim() == 4 else alpha_image
+        return IO.NodeOutput(render, alpha_mask)
+
+
+class BeebleExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [
+            BeebleSwitchXVideoEdit,
+            BeebleSwitchXImageEdit,
+        ]
+
+
+async def comfy_entrypoint() -> BeebleExtension:
+    return BeebleExtension()

comfy_api_nodes/util/client.py

@@ -86,7 +86,7 @@ class _PollUIState:
 _RETRY_STATUS = {408, 500, 502, 503, 504}  # status 429 is handled separately
 COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed", "finished", "done", "complete"]
 FAILED_STATUSES = ["cancelled", "canceled", "canceling", "fail", "failed", "error"]
-QUEUED_STATUSES = ["created", "queued", "queueing", "submitted", "initializing", "wait"]
+QUEUED_STATUSES = ["created", "queued", "queueing", "submitted", "initializing", "wait", "in_queue"]
 
 
 async def sync_op(

comfy_extras/nodes_load_3d.py

@@ -47,6 +47,7 @@ class Load3D(IO.ComfyNode):
                 IO.Load3DCamera.Output(display_name="camera_info"),
                 IO.Video.Output(display_name="recording_video"),
                 IO.File3DAny.Output(display_name="model_3d"),
+                IO.Load3DModelInfo.Output(display_name="model_info"),
             ],
         )
 
@@ -69,7 +70,8 @@ class Load3D(IO.ComfyNode):
             video = InputImpl.VideoFromFile(recording_video_path)
 
         file_3d = Types.File3D(folder_paths.get_annotated_filepath(model_file))
-        return IO.NodeOutput(output_image, output_mask, model_file, normal_image, image['camera_info'], video, file_3d)
+        model_info = image.get('model_info', [])
+        return IO.NodeOutput(output_image, output_mask, model_file, normal_image, image['camera_info'], video, file_3d, model_info)
 
     process = execute  # TODO: remove
 

comfy_extras/nodes_lt_upsampler.py

@@ -1,32 +1,32 @@
 from comfy import model_management
+from comfy_api.latest import ComfyExtension, IO
+from typing_extensions import override
 import math
 
-class LTXVLatentUpsampler:
+
+class LTXVLatentUpsampler(IO.ComfyNode):
     """
     Upsamples a video latent by a factor of 2.
     """
 
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "samples": ("LATENT",),
-                "upscale_model": ("LATENT_UPSCALE_MODEL",),
-                "vae": ("VAE",),
-            }
-        }
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="LTXVLatentUpsampler",
+            category="latent/video",
+            is_experimental=True,
+            inputs=[
+                IO.Latent.Input("samples"),
+                IO.LatentUpscaleModel.Input("upscale_model"),
+                IO.Vae.Input("vae"),
+            ],
+            outputs=[
+                IO.Latent.Output(),
+            ],
+        )
 
-    RETURN_TYPES = ("LATENT",)
-    FUNCTION = "upsample_latent"
-    CATEGORY = "latent/video"
-    EXPERIMENTAL = True
-
-    def upsample_latent(
-        self,
-        samples: dict,
-        upscale_model,
-        vae,
-    ) -> tuple:
+    @classmethod
+    def execute(cls, samples, upscale_model, vae) -> IO.NodeOutput:
         """
         Upsample the input latent using the provided model.
 
@@ -34,7 +34,6 @@ class LTXVLatentUpsampler:
             samples (dict): Input latent samples
             upscale_model (LatentUpsampler): Loaded upscale model
             vae: VAE model for normalization
-            auto_tiling (bool): Whether to automatically tile the input for processing
 
         Returns:
             tuple: Tuple containing the upsampled latent
@@ -67,9 +66,16 @@ class LTXVLatentUpsampler:
         return_dict = samples.copy()
         return_dict["samples"] = upsampled_latents
         return_dict.pop("noise_mask", None)
-        return (return_dict,)
+        return IO.NodeOutput(return_dict)
+
+    upsample_latent = execute  # TODO: remove
 
 
-NODE_CLASS_MAPPINGS = {
-    "LTXVLatentUpsampler": LTXVLatentUpsampler,
-}
+class LTXVLatentUpsamplerExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [LTXVLatentUpsampler]
+
+
+async def comfy_entrypoint() -> LTXVLatentUpsamplerExtension:
+    return LTXVLatentUpsamplerExtension()