Merge 8eee52de37 into f3ea976cba

comfy/cli_args.py

@@ -91,6 +91,7 @@ parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE"
 
 parser.add_argument("--oneapi-device-selector", type=str, default=None, metavar="SELECTOR_STRING", help="Sets the oneAPI device(s) this instance will use.")
 parser.add_argument("--supports-fp8-compute", action="store_true", help="ComfyUI will act like if the device supports fp8 compute.")
+parser.add_argument("--enable-triton-backend", action="store_true", help="ComfyUI will enable the use of Triton backend in comfy-kitchen. Is disabled at launch by default.")
 
 class LatentPreviewMethod(enum.Enum):
     NoPreviews = "none"

comfy/context_windows.py

@@ -63,7 +63,11 @@ class IndexListContextWindow(ContextWindowABC):
             dim = self.dim
         if dim == 0 and full.shape[dim] == 1:
             return full
-        idx = tuple([slice(None)] * dim + [self.index_list])
+        indices = self.index_list
+        anchor_idx = getattr(self, 'causal_anchor_index', None)
+        if anchor_idx is not None and anchor_idx >= 0:
+            indices = [anchor_idx] + list(indices)
+        idx = tuple([slice(None)] * dim + [indices])
         window = full[idx]
         if retain_index_list:
             idx = tuple([slice(None)] * dim + [retain_index_list])
@@ -113,7 +117,14 @@ def slice_cond(cond_value, window: IndexListContextWindow, x_in: torch.Tensor, d
 
     # skip leading latent positions that have no corresponding conditioning (e.g. reference frames)
     if temporal_offset > 0:
-        indices = [i - temporal_offset for i in window.index_list[temporal_offset:]]
+        anchor_idx = getattr(window, 'causal_anchor_index', None)
+        if anchor_idx is not None and anchor_idx >= 0:
+            # anchor occupies one of the no-cond positions, so skip one fewer from window.index_list
+            skip_count = temporal_offset - 1
+        else:
+            skip_count = temporal_offset
+
+        indices = [i - temporal_offset for i in window.index_list[skip_count:]]
         indices = [i for i in indices if 0 <= i]
     else:
         indices = list(window.index_list)
@@ -150,7 +161,8 @@ class ContextFuseMethod:
 ContextResults = collections.namedtuple("ContextResults", ['window_idx', 'sub_conds_out', 'sub_conds', 'window'])
 class IndexListContextHandler(ContextHandlerABC):
     def __init__(self, context_schedule: ContextSchedule, fuse_method: ContextFuseMethod, context_length: int=1, context_overlap: int=0, context_stride: int=1,
-                 closed_loop: bool=False, dim:int=0, freenoise: bool=False, cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False):
+                 closed_loop: bool=False, dim:int=0, freenoise: bool=False, cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False,
+                 causal_window_fix: bool=True):
         self.context_schedule = context_schedule
         self.fuse_method = fuse_method
         self.context_length = context_length
@@ -162,6 +174,7 @@ class IndexListContextHandler(ContextHandlerABC):
         self.freenoise = freenoise
         self.cond_retain_index_list = [int(x.strip()) for x in cond_retain_index_list.split(",")] if cond_retain_index_list else []
         self.split_conds_to_windows = split_conds_to_windows
+        self.causal_window_fix = causal_window_fix
 
         self.callbacks = {}
 
@@ -318,6 +331,14 @@ class IndexListContextHandler(ContextHandlerABC):
             # allow processing to end between context window executions for faster Cancel
             comfy.model_management.throw_exception_if_processing_interrupted()
 
+            # causal_window_fix: prepend a pre-window frame that will be stripped post-forward
+            anchor_applied = False
+            if self.causal_window_fix:
+                anchor_idx = window.index_list[0] - 1
+                if 0 <= anchor_idx < x_in.size(self.dim):
+                    window.causal_anchor_index = anchor_idx
+                    anchor_applied = True
+
             for callback in comfy.patcher_extension.get_all_callbacks(IndexListCallbacks.EVALUATE_CONTEXT_WINDOWS, self.callbacks):
                 callback(self, model, x_in, conds, timestep, model_options, window_idx, window, model_options, device, first_device)
 
@@ -332,6 +353,12 @@ class IndexListContextHandler(ContextHandlerABC):
             if device is not None:
                 for i in range(len(sub_conds_out)):
                     sub_conds_out[i] = sub_conds_out[i].to(x_in.device)
+
+            # strip causal_window_fix anchor if applied
+            if anchor_applied:
+                for i in range(len(sub_conds_out)):
+                    sub_conds_out[i] = sub_conds_out[i].narrow(self.dim, 1, sub_conds_out[i].shape[self.dim] - 1)
+
             results.append(ContextResults(window_idx, sub_conds_out, sub_conds, window))
         return results
 

comfy/quant_ops.py

@@ -1,6 +1,8 @@
 import torch
 import logging
 
+from comfy.cli_args import args
+
 try:
     import comfy_kitchen as ck
     from comfy_kitchen.tensor import (
@@ -21,7 +23,15 @@ try:
             ck.registry.disable("cuda")
             logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
 
-    ck.registry.disable("triton")
+    if args.enable_triton_backend:
+        try:
+            import triton
+            logging.info("Found triton %s. Enabling comfy-kitchen triton backend.", triton.__version__)
+        except ImportError as e:
+            logging.error(f"Failed to import triton, Error: {e}, the comfy-kitchen triton backend will not be available.")
+            ck.registry.disable("triton")
+    else:
+        ck.registry.disable("triton")
     for k, v in ck.list_backends().items():
         logging.info(f"Found comfy_kitchen backend {k}: {v}")
 except ImportError as e:

comfy_extras/nodes_context_windows.py

@@ -29,6 +29,7 @@ class ContextWindowsManualNode(io.ComfyNode):
                 io.Boolean.Input("freenoise", default=False, tooltip="Whether to apply FreeNoise noise shuffling, improves window blending."),
                 io.String.Input("cond_retain_index_list", default="", tooltip="List of latent indices to retain in the conditioning tensors for each window, for example setting this to '0' will use the initial start image for each window."),
                 io.Boolean.Input("split_conds_to_windows", default=False, tooltip="Whether to split multiple conditionings (created by ConditionCombine) to each window based on region index."),
+                io.Boolean.Input("causal_window_fix", default=True, tooltip="Whether to add a causal fix frame to non-0-indexed context windows."),
             ],
             outputs=[
                 io.Model.Output(tooltip="The model with context windows applied during sampling."),
@@ -38,7 +39,7 @@ class ContextWindowsManualNode(io.ComfyNode):
 
     @classmethod
     def execute(cls, model: io.Model.Type, context_length: int, context_overlap: int, context_schedule: str, context_stride: int, closed_loop: bool, fuse_method: str, dim: int, freenoise: bool,
-                cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False) -> io.Model:
+                cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False, causal_window_fix: bool=True) -> io.Model:
         model = model.clone()
         model.model_options["context_handler"] = comfy.context_windows.IndexListContextHandler(
             context_schedule=comfy.context_windows.get_matching_context_schedule(context_schedule),
@@ -50,7 +51,8 @@ class ContextWindowsManualNode(io.ComfyNode):
             dim=dim,
             freenoise=freenoise,
             cond_retain_index_list=cond_retain_index_list,
-            split_conds_to_windows=split_conds_to_windows
+            split_conds_to_windows=split_conds_to_windows,
+            causal_window_fix=causal_window_fix,
         )
         # make memory usage calculation only take into account the context window latents
         comfy.context_windows.create_prepare_sampling_wrapper(model)

comfy_extras/nodes_post_processing.py

@@ -666,12 +666,13 @@ class ColorTransfer(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="ColorTransfer",
+            display_name="Color Transfer",
             category="image/postprocessing",
             description="Match the colors of one image to another using various algorithms.",
             search_aliases=["color match", "color grading", "color correction", "match colors", "color transform", "mkl", "reinhard", "histogram"],
             inputs=[
                 io.Image.Input("image_target", tooltip="Image(s) to apply the color transform to."),
-                io.Image.Input("image_ref", optional=True, tooltip="Reference image(s) to match colors to. If not provided, processing is skipped"),
+                io.Image.Input("image_ref", tooltip="Reference image(s) to match colors to."),
                 io.Combo.Input("method", options=['reinhard_lab', 'mkl_lab', 'histogram'],),
                 io.DynamicCombo.Input("source_stats",
                     tooltip="per_frame: each frame matched to image_ref individually. uniform: pool stats across all source frames as baseline, match to image_ref. target_frame: use one chosen frame as the baseline for the transform to image_ref, applied uniformly to all frames (preserves relative differences)",

comfy_extras/nodes_primitive.py

@@ -49,7 +49,7 @@ class Int(io.ComfyNode):
             display_name="Int",
             category="utils/primitive",
             inputs=[
-                io.Int.Input("value", min=-sys.maxsize, max=sys.maxsize, control_after_generate=True),
+                io.Int.Input("value", min=-sys.maxsize, max=sys.maxsize, control_after_generate=io.ControlAfterGenerate.fixed),
             ],
             outputs=[io.Int.Output()],
         )

extra_model_paths.yaml.example

@@ -28,7 +28,7 @@
 #config for a1111 ui
 #all you have to do is uncomment this (remove the #) and change the base_path to where yours is installed
 
-#a111:
+#a1111:
 #     base_path: path/to/stable-diffusion-webui/
 #     checkpoints: models/Stable-diffusion
 #     configs: models/Stable-diffusion

nodes.py

@@ -1754,57 +1754,49 @@ class LoadImage:
 
         return True
 
-class LoadImageMask:
+
+class LoadImageMask(LoadImage):
     ESSENTIALS_CATEGORY = "Image Tools"
     SEARCH_ALIASES = ["import mask", "alpha mask", "channel mask"]
 
     _color_channels = ["alpha", "red", "green", "blue"]
+
     @classmethod
     def INPUT_TYPES(s):
-        input_dir = folder_paths.get_input_directory()
-        files = [f for f in os.listdir(input_dir) if os.path.isfile(os.path.join(input_dir, f))]
-        return {"required":
-                    {"image": (sorted(files), {"image_upload": True}),
-                     "channel": (s._color_channels, ), }
-                }
+        types = super().INPUT_TYPES()
+        return {
+            "required": {
+                **types["required"],
+                "channel": (s._color_channels, )
+            }
+        }
 
     CATEGORY = "mask"
-
     RETURN_TYPES = ("MASK",)
-    FUNCTION = "load_image"
-    def load_image(self, image, channel):
-        image_path = folder_paths.get_annotated_filepath(image)
-        i = node_helpers.pillow(Image.open, image_path)
-        i = node_helpers.pillow(ImageOps.exif_transpose, i)
-        if i.getbands() != ("R", "G", "B", "A"):
-            if i.mode == 'I':
-                i = i.point(lambda i: i * (1 / 255))
-            i = i.convert("RGBA")
-        mask = None
+    FUNCTION = "load_image_mask"
+
+    def load_image_mask(self, image, channel):
+        image_tensor, mask_tensor = super().load_image(image)
         c = channel[0].upper()
-        if c in i.getbands():
-            mask = np.array(i.getchannel(c)).astype(np.float32) / 255.0
-            mask = torch.from_numpy(mask)
-            if c == 'A':
-                mask = 1. - mask
+
+        if c == 'A':
+            return (mask_tensor,)
+
+        channel_idx = {'R': 0, 'G': 1, 'B': 2}.get(c, 0)
+
+        if channel_idx < image_tensor.shape[-1]:
+            return (image_tensor[..., channel_idx].clone(),)
         else:
-            mask = torch.zeros((64,64), dtype=torch.float32, device="cpu")
-        return (mask.unsqueeze(0),)
+            empty_mask = torch.zeros(
+                image_tensor.shape[:-1],
+                dtype=image_tensor.dtype,
+                device=image_tensor.device
+            )
+            return (empty_mask,)
 
     @classmethod
     def IS_CHANGED(s, image, channel):
-        image_path = folder_paths.get_annotated_filepath(image)
-        m = hashlib.sha256()
-        with open(image_path, 'rb') as f:
-            m.update(f.read())
-        return m.digest().hex()
-
-    @classmethod
-    def VALIDATE_INPUTS(s, image):
-        if not folder_paths.exists_annotated_filepath(image):
-            return "Invalid image file: {}".format(image)
-
-        return True
+        return super().IS_CHANGED(image)
 
 
 class LoadImageOutput(LoadImage):