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_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