comfy-diffusion 2.3.0__tar.gz → 2.4.0__tar.gz

{comfy_diffusion-2.3.0 → comfy_diffusion-2.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: comfy-diffusion
-Version: 2.3.0
+Version: 2.4.0
 Summary: ComfyUI inference engine as a standalone Python library (no server, no UI).
 Author: Alejandro Quintero
 License: GPL-3.0-only
@@ -130,7 +130,7 @@ from comfy_diffusion.sampling import sample
 
 ## Quick Start
 
-Call `check_runtime()` before loading models or sampling. On first runtime use, comfy-diffusion can perform an automatic download of the pinned ComfyUI release when the vendored runtime is missing. Expected failures are returned as an error dict instead of being raised; `check_runtime()` returns an error dict for runtime bootstrap problems.
+Call `check_runtime()` before loading models or sampling. On first runtime use, comfy-diffusion can perform an automatic download of the pinned ComfyUI ref when the vendored runtime is missing. Expected failures are returned as an error dict instead of being raised; `check_runtime()` returns an error dict for runtime bootstrap problems.
 
 Example:
 

{comfy_diffusion-2.3.0 → comfy_diffusion-2.4.0}/README.md

@@ -52,7 +52,7 @@ from comfy_diffusion.sampling import sample
 
 ## Quick Start
 
-Call `check_runtime()` before loading models or sampling. On first runtime use, comfy-diffusion can perform an automatic download of the pinned ComfyUI release when the vendored runtime is missing. Expected failures are returned as an error dict instead of being raised; `check_runtime()` returns an error dict for runtime bootstrap problems.
+Call `check_runtime()` before loading models or sampling. On first runtime use, comfy-diffusion can perform an automatic download of the pinned ComfyUI ref when the vendored runtime is missing. Expected failures are returned as an error dict instead of being raised; `check_runtime()` returns an error dict for runtime bootstrap problems.
 
 Example:
 

{comfy_diffusion-2.3.0 → comfy_diffusion-2.4.0}/comfy_diffusion/_runtime.py

@@ -13,10 +13,10 @@ import urllib.request
 import zipfile
 from pathlib import Path
 
-COMFYUI_PINNED_TAG = "v0.24.1"
+COMFYUI_PINNED_REF = "fb991e2c1e7476809d566a4620c2132e05a466dd"
 COMFYUI_PINNED_ARCHIVE_URL = (
-    "https://github.com/Comfy-Org/ComfyUI/archive/refs/tags/"
-    f"{COMFYUI_PINNED_TAG}.zip"
+    "https://github.com/Comfy-Org/ComfyUI/archive/"
+    f"{COMFYUI_PINNED_REF}.zip"
 )
 
 
@@ -40,7 +40,7 @@ def _has_comfyui_runtime(comfyui_root: Path) -> bool:
 
 
 def _download_and_extract_pinned_comfyui(comfyui_root: Path) -> None:
-    """Download and extract the pinned ComfyUI release into vendor/ComfyUI."""
+    """Download and extract the pinned ComfyUI ref into vendor/ComfyUI."""
     vendor_dir = comfyui_root.parent
     vendor_dir.mkdir(parents=True, exist_ok=True)
 
@@ -67,7 +67,7 @@ def _download_and_extract_pinned_comfyui(comfyui_root: Path) -> None:
 
 
 def ensure_comfyui_available() -> Path:
-    """Ensure vendored ComfyUI exists; download pinned release if missing."""
+    """Ensure vendored ComfyUI exists; download pinned ref if missing."""
     comfyui_root = _comfyui_root()
 
     if not _has_comfyui_runtime(comfyui_root):

{comfy_diffusion-2.3.0 → comfy_diffusion-2.4.0}/comfy_diffusion/cli/main.py

@@ -10,7 +10,7 @@ from typing import Annotated, Any
 import click
 import typer
 
-from comfy_diffusion._runtime import COMFYUI_PINNED_TAG, _comfyui_root
+from comfy_diffusion._runtime import COMFYUI_PINNED_REF, _comfyui_root
 from comfy_diffusion.downloader import (
     CivitAIModelEntry,
     HFModelEntry,
@@ -116,7 +116,7 @@ def runtime_paths(
         "package_root": str(_package_root()),
         "comfyui_root": str(_comfyui_root()),
         "models_dir": str(resolved_models_dir),
-        "comfyui_pinned_tag": COMFYUI_PINNED_TAG,
+        "comfyui_pinned_ref": COMFYUI_PINNED_REF,
     }
     if json_output:
         _echo_json(payload)

{comfy_diffusion-2.3.0 → comfy_diffusion-2.4.0}/comfy_diffusion/conditioning.py

@@ -181,6 +181,83 @@ def wan_image_to_video(
     return positive, negative, {"samples": latent}
 
 
+def _resize_long_edge(image: Any, max_size: int, stride: int = 16) -> Any:
+    """Resize an image batch so the long edge is capped and dimensions snap to stride."""
+    _, _, comfy_utils, _ = _get_wan_conditioning_dependencies()
+
+    height = image.shape[1]
+    width = image.shape[2]
+    scale = min(max_size / max(height, width), 1.0)
+    resized_height = max(stride, round(height * scale / stride) * stride)
+    resized_width = max(stride, round(width * scale / stride) * stride)
+    return comfy_utils.common_upscale(
+        image[:, :, :, :3].movedim(-1, 1),
+        resized_width,
+        resized_height,
+        "area",
+        "disabled",
+    ).movedim(1, -1)
+
+
+def bernini_conditioning(
+    positive: Any,
+    negative: Any,
+    vae: _VaeEncoder,
+    width: int = 832,
+    height: int = 480,
+    length: int = 81,
+    batch_size: int = 1,
+    *,
+    source_video: Any | None = None,
+    reference_video: Any | None = None,
+    reference_images: Sequence[Any] | None = None,
+    ref_max_size: int = 848,
+) -> tuple[Any, Any, dict[str, Any]]:
+    """Build Bernini in-context WAN conditioning and an empty video latent.
+
+    Mirrors ComfyUI's ``BerniniConditioning`` node.  Context streams are VAE
+    encoded in this order: source video, reference video, then each reference
+    image.  When any context stream is present, the resulting latent list is
+    attached to both positive and negative conditioning as ``context_latents``.
+    """
+    torch, model_management, comfy_utils, node_helpers = _get_wan_conditioning_dependencies()
+
+    latent = torch.zeros(
+        [batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8],
+        device=model_management.intermediate_device(),
+    )
+
+    context = []
+    if source_video is not None:
+        video = comfy_utils.common_upscale(
+            source_video[:length, :, :, :3].movedim(-1, 1),
+            width,
+            height,
+            "area",
+            "center",
+        ).movedim(1, -1)
+        context.append(vae.encode(video[:, :, :, :3]))
+
+    if reference_video is not None:
+        resized_reference_video = _resize_long_edge(reference_video[:length], ref_max_size)
+        context.append(vae.encode(resized_reference_video[:, :, :, :3]))
+
+    if reference_images:
+        for images in reference_images:
+            if images is None:
+                continue
+            for index in range(images.shape[0]):
+                image = _resize_long_edge(images[index : index + 1], ref_max_size)
+                context.append(vae.encode(image[:, :, :, :3]))
+
+    if context:
+        values = {"context_latents": context}
+        positive = node_helpers.conditioning_set_values(positive, values)
+        negative = node_helpers.conditioning_set_values(negative, values)
+
+    return positive, negative, {"samples": latent}
+
+
 def wan_first_last_frame_to_video(
     positive: Any,
     negative: Any,
@@ -365,6 +442,71 @@ def ltxv_crop_guides(
     return positive, negative, {"samples": latent_image, "noise_mask": noise_mask}
 
 
+def _get_void_quadmask_preprocess_type() -> Any:
+    """Resolve ComfyUI VOIDQuadmaskPreprocess node at call time."""
+    from ._runtime import ensure_comfyui_on_path
+
+    ensure_comfyui_on_path()
+    from comfy_extras.nodes_void import VOIDQuadmaskPreprocess
+
+    return VOIDQuadmaskPreprocess
+
+
+def _get_void_inpaint_conditioning_type() -> Any:
+    """Resolve ComfyUI VOIDInpaintConditioning node at call time."""
+    from ._runtime import ensure_comfyui_on_path
+
+    ensure_comfyui_on_path()
+    from comfy_extras.nodes_void import VOIDInpaintConditioning
+
+    return VOIDInpaintConditioning
+
+
+def void_quadmask_preprocess(mask: Any, dilate_width: int = 0) -> Any:
+    """Preprocess a VOID quadmask video mask.
+
+    Mirrors ComfyUI's ``VOIDQuadmaskPreprocess``.  The mask is quantized to the
+    four semantic VOID levels, inverted, and normalized for
+    ``void_inpaint_conditioning``.
+    """
+    node_type = _get_void_quadmask_preprocess_type()
+    output = node_type.execute(mask=mask, dilate_width=dilate_width)
+    result = getattr(output, "result", output)
+    return result[0]
+
+
+def void_inpaint_conditioning(
+    positive: Any,
+    negative: Any,
+    vae: Any,
+    video: Any,
+    quadmask: Any,
+    width: int = 672,
+    height: int = 384,
+    length: int = 45,
+    batch_size: int = 1,
+) -> tuple[Any, Any, dict[str, Any]]:
+    """Build VOID inpainting conditioning and the matching noise latent.
+
+    Mirrors ComfyUI's ``VOIDInpaintConditioning`` for CogVideoX-Fun-V1.5 VOID
+    inpainting models.
+    """
+    node_type = _get_void_inpaint_conditioning_type()
+    output = node_type.execute(
+        positive=positive,
+        negative=negative,
+        vae=vae,
+        video=video,
+        quadmask=quadmask,
+        width=width,
+        height=height,
+        length=length,
+        batch_size=batch_size,
+    )
+    result = getattr(output, "result", output)
+    return result[0], result[1], result[2]
+
+
 def _get_ltxv_add_guide_type() -> Any:
     """Resolve ComfyUI LTXVAddGuide node at call time."""
     from comfy_extras.nodes_lt import LTXVAddGuide
@@ -2164,6 +2306,7 @@ __all__ = [
     "encode_prompt_flux",
     "encode_clip_vision",
     "wan_image_to_video",
+    "bernini_conditioning",
     "wan_first_last_frame_to_video",
     "wan_vace_to_video",
     "wan_fun_control_to_video",
@@ -2183,6 +2326,8 @@ __all__ = [
     "ltxv_img_to_video",
     "ltxv_conditioning",
     "ltxv_crop_guides",
+    "void_quadmask_preprocess",
+    "void_inpaint_conditioning",
     "ltx_add_video_ic_lora_guide",
     "conditioning_zero_out",
     "conditioning_combine",

{comfy_diffusion-2.3.0 → comfy_diffusion-2.4.0}/comfy_diffusion/image.py

@@ -324,6 +324,15 @@ def image_to_tensor(image: PILImage.Image) -> Any:
     return torch.tensor([rows], dtype=torch.float32)
 
 
+def images_to_tensor(images: list[PILImage.Image]) -> Any:
+    """Convert PIL images to a BHWC float32 tensor."""
+    if not images:
+        raise ValueError("images must not be empty")
+    torch = _get_torch_module()
+    rows = [_pixels_to_float_rows(image.convert("RGB"), channels=3) for image in images]
+    return torch.tensor(rows, dtype=torch.float32)
+
+
 def ltxv_preprocess(image: Any, width: int, height: int, img_compression: int = 35) -> Any:
     """Preprocess image batch for LTXV img2vid with center resize and node compression."""
     comfy_utils, ltxv_preprocess_type = _get_ltxv_preprocess_dependencies()
@@ -788,6 +797,7 @@ def flux_kontext_image_scale(image: Any) -> Any:
 __all__ = [
     "load_image",
     "image_to_tensor",
+    "images_to_tensor",
     "image_pad_for_outpaint",
     "image_upscale_with_model",
     "image_from_batch",

{comfy_diffusion-2.3.0 → comfy_diffusion-2.4.0}/comfy_diffusion/latent.py

@@ -93,6 +93,7 @@ def _get_latent_concat_type() -> Any:
     from ._runtime import ensure_comfyui_on_path
 
     ensure_comfyui_on_path()
+    import comfy_extras.nodes_post_processing  # noqa: F401
     from comfy_extras.nodes_latent import LatentConcat
 
     return LatentConcat
@@ -103,6 +104,7 @@ def _get_latent_cut_type() -> Any:
     from ._runtime import ensure_comfyui_on_path
 
     ensure_comfyui_on_path()
+    import comfy_extras.nodes_post_processing  # noqa: F401
     from comfy_extras.nodes_latent import LatentCut
 
     return LatentCut
@@ -113,6 +115,7 @@ def _get_replace_video_latent_frames_type() -> Any:
     from ._runtime import ensure_comfyui_on_path
 
     ensure_comfyui_on_path()
+    import comfy_extras.nodes_post_processing  # noqa: F401
     from comfy_extras.nodes_latent import ReplaceVideoLatentFrames
 
     return ReplaceVideoLatentFrames

{comfy_diffusion-2.3.0 → comfy_diffusion-2.4.0}/comfy_diffusion/models.py

@@ -76,6 +76,9 @@ class ModelManager:
         folder_paths.add_model_folder_path(
             "audio_encoders", str(self.models_dir / "audio_encoders"), is_default=True
         )
+        folder_paths.add_model_folder_path(
+            "optical_flow", str(self.models_dir / "optical_flow"), is_default=True
+        )
 
     def load_checkpoint(self, filename: str) -> CheckpointResult:
         """Load a checkpoint by filename from the configured checkpoints directory."""
@@ -409,6 +412,29 @@ class ModelManager:
             )
         return audio_encoder
 
+    def load_optical_flow(self, path: str | Path) -> Any:
+        """Load an optical-flow model from a path or filename.
+
+        Mirrors ComfyUI's ``OpticalFlowLoader`` used by VOID.  If ``path`` is an
+        absolute path, the file must exist.  Otherwise it is treated as a
+        filename under ``models_dir/optical_flow``.
+        """
+        ensure_comfyui_on_path()
+
+        p = Path(path)
+        if p.is_absolute():
+            if not p.is_file():
+                raise FileNotFoundError(f"optical flow model file not found: {p}")
+            model_name = p.name
+        else:
+            model_name = path if isinstance(path, str) else p.name
+
+        from comfy_extras.nodes_void import OpticalFlowLoader
+
+        output = OpticalFlowLoader.execute(model_name)
+        result = getattr(output, "result", output)
+        return result[0]
+
     def load_vae_kj(
         self,
         path: str | Path,