[pull] master from Comfy-Org:master

app/frontend_management.py

@@ -27,7 +27,7 @@ def frontend_install_warning_message():
     return f"""
 {get_missing_requirements_message()}
 
-This error is happening because the ComfyUI frontend is no longer shipped as part of the main repo but as a pip package instead.
+The ComfyUI frontend is shipped in a pip package so it needs to be updated separately from the ComfyUI code.
 """.strip()
 
 def parse_version(version: str) -> tuple[int, int, int]:

app/node_replace_manager.py

@@ -1,5 +1,7 @@
 from __future__ import annotations
 
+import logging
+
 from aiohttp import web
 
 from typing import TYPE_CHECKING, TypedDict
@@ -31,8 +33,22 @@ def __init__(self):
         self._replacements: dict[str, list[NodeReplace]] = {}
 
     def register(self, node_replace: NodeReplace):
-        """Register a node replacement mapping."""
-        self._replacements.setdefault(node_replace.old_node_id, []).append(node_replace)
+        """Register a node replacement mapping.
+
+        Idempotent: if a replacement with the same (old_node_id, new_node_id)
+        is already registered, the duplicate is ignored. This prevents stale
+        entries from accumulating when custom nodes are reloaded in the same
+        process (e.g. via ComfyUI-Manager).
+        """
+        existing = self._replacements.setdefault(node_replace.old_node_id, [])
+        for entry in existing:
+            if entry.new_node_id == node_replace.new_node_id:
+                logging.debug(
+                    "Node replacement %s -> %s already registered, ignoring duplicate.",
+                    node_replace.old_node_id, node_replace.new_node_id,
+                )
+                return
+        existing.append(node_replace)
 
     def get_replacement(self, old_node_id: str) -> list[NodeReplace] | None:
         """Get replacements for an old node ID."""

comfy/k_diffusion/sampling.py

@@ -1859,6 +1859,23 @@ def sample_ar_video(model, x, sigmas, extra_args=None, callback=None, disable=No
     output = torch.zeros_like(x)
     s_in = x.new_ones([x.shape[0]])
     current_start_frame = 0
+
+    # I2V: seed KV cache with the initial image latent before the denoising loop
+    initial_latent = transformer_options.get("ar_config", {}).get("initial_latent", None)
+    if initial_latent is not None:
+        initial_latent = inner_model.process_latent_in(initial_latent).to(device=device, dtype=model_dtype)
+        n_init = initial_latent.shape[2]
+        output[:, :, :n_init] = initial_latent
+
+        ar_state = {"start_frame": 0, "kv_caches": kv_caches, "crossattn_caches": crossattn_caches}
+        transformer_options["ar_state"] = ar_state
+        zero_sigma = sigmas.new_zeros([1])
+        _ = model(initial_latent, zero_sigma * s_in, **extra_args)
+
+        current_start_frame = n_init
+        remaining = lat_t - n_init
+        num_blocks = -(-remaining // num_frame_per_block)
+
     num_sigma_steps = len(sigmas) - 1
     total_real_steps = num_blocks * num_sigma_steps
     step_count = 0

comfy/ldm/sam3/detector.py

@@ -561,7 +561,8 @@ def forward_segment(self, images, point_inputs=None, box_inputs=None, mask_input
         return high_res_masks
 
     def forward_video(self, images, initial_masks, pbar=None, text_prompts=None,
-                       new_det_thresh=0.5, max_objects=0, detect_interval=1):
+                       new_det_thresh=0.5, max_objects=0, detect_interval=1,
+                       target_device=None, target_dtype=None):
         """Track video with optional per-frame text-prompted detection."""
         bb = self.detector.backbone["vision_backbone"]
 
@@ -589,8 +590,10 @@ def detect_fn(trunk_out):
             return self.tracker.track_video_with_detection(
                 backbone_fn, images, initial_masks, detect_fn,
                 new_det_thresh=new_det_thresh, max_objects=max_objects,
-                detect_interval=detect_interval, backbone_obj=bb, pbar=pbar)
+                detect_interval=detect_interval, backbone_obj=bb, pbar=pbar,
+                target_device=target_device, target_dtype=target_dtype)
         # SAM3 (non-multiplex) — no detection support, requires initial masks
         if initial_masks is None:
             raise ValueError("SAM3 (non-multiplex) requires initial_mask for video tracking")
-        return self.tracker.track_video(backbone_fn, images, initial_masks, pbar=pbar, backbone_obj=bb)
+        return self.tracker.track_video(backbone_fn, images, initial_masks, pbar=pbar, backbone_obj=bb,
+                                         target_device=target_device, target_dtype=target_dtype)

comfy/ldm/sam3/tracker.py

@@ -200,8 +200,13 @@ def pack_masks(masks):
 
 def unpack_masks(packed):
     """Unpack bit-packed [*, H, W//8] uint8 to bool [*, H, W*8]."""
-    shifts = torch.arange(8, device=packed.device)
-    return ((packed.unsqueeze(-1) >> shifts) & 1).view(*packed.shape[:-1], -1).bool()
+    bits = torch.tensor([1, 2, 4, 8, 16, 32, 64, 128], dtype=torch.uint8, device=packed.device)
+    return (packed.unsqueeze(-1) & bits).bool().view(*packed.shape[:-1], -1)
+
+
+def _prep_frame(images, idx, device, dt, size):
+    """Slice CPU full-res frames, transfer to GPU in target dtype, and resize to (size, size)."""
+    return comfy.utils.common_upscale(images[idx].to(device=device, dtype=dt), size, size, "bicubic", crop="disabled")
 
 
 def _compute_backbone(backbone_fn, frame, frame_idx=None):
@@ -1078,16 +1083,19 @@ def _compute_backbone_frame(self, backbone_fn, frame, frame_idx=None):
         # SAM3: drop last FPN level
         return vision_feats[:-1], vision_pos[:-1], feat_sizes[:-1]
 
-    def _track_single_object(self, backbone_fn, images, initial_mask, pbar=None):
+    def _track_single_object(self, backbone_fn, images, initial_mask, pbar=None,
+                             target_device=None, target_dtype=None):
         """Track one object, computing backbone per frame to save VRAM."""
         N = images.shape[0]
-        device, dt = images.device, images.dtype
+        device = target_device if target_device is not None else images.device
+        dt = target_dtype if target_dtype is not None else images.dtype
+        size = self.image_size
         output_dict = {"cond_frame_outputs": {}, "non_cond_frame_outputs": {}}
         all_masks = []
 
         for frame_idx in tqdm(range(N), desc="tracking"):
             vision_feats, vision_pos, feat_sizes = self._compute_backbone_frame(
-                backbone_fn, images[frame_idx:frame_idx + 1], frame_idx=frame_idx)
+                backbone_fn, _prep_frame(images, slice(frame_idx, frame_idx + 1), device, dt, size), frame_idx=frame_idx)
             mask_input = None
             if frame_idx == 0:
                 mask_input = F.interpolate(initial_mask.to(device=device, dtype=dt),
@@ -1114,12 +1122,13 @@ def _track_single_object(self, backbone_fn, images, initial_mask, pbar=None):
 
         return torch.cat(all_masks, dim=0)  # [N, 1, H, W]
 
-    def track_video(self, backbone_fn, images, initial_masks, pbar=None, **kwargs):
+    def track_video(self, backbone_fn, images, initial_masks, pbar=None,
+                    target_device=None, target_dtype=None, **kwargs):
         """Track one or more objects across video frames.
 
         Args:
             backbone_fn: callable that returns (sam2_features, sam2_positions, trunk_out) for a frame
-            images: [N, 3, 1008, 1008] video frames
+            images: [N, 3, H, W] CPU full-res video frames (resized per-frame to self.image_size)
             initial_masks: [N_obj, 1, H, W] binary masks for first frame (one per object)
             pbar: optional progress bar
 
@@ -1130,7 +1139,8 @@ def track_video(self, backbone_fn, images, initial_masks, pbar=None, **kwargs):
         per_object = []
         for obj_idx in range(N_obj):
             obj_masks = self._track_single_object(
-                backbone_fn, images, initial_masks[obj_idx:obj_idx + 1], pbar=pbar)
+                backbone_fn, images, initial_masks[obj_idx:obj_idx + 1], pbar=pbar,
+                target_device=target_device, target_dtype=target_dtype)
             per_object.append(obj_masks)
 
         return torch.cat(per_object, dim=1)  # [N, N_obj, H, W]
@@ -1632,11 +1642,18 @@ def _match_and_add_detections(self, det_masks, det_scores, current_out, mux_stat
             return det_scores[new_dets].tolist() if det_scores is not None else [0.0] * new_dets.sum().item()
         return []
 
+    INTERNAL_MAX_OBJECTS = 64  # Hard ceiling on accumulated tracks; max_objects=0 or any value above this is clamped here.
+
     def track_video_with_detection(self, backbone_fn, images, initial_masks, detect_fn=None,
                                    new_det_thresh=0.5, max_objects=0, detect_interval=1,
-                                   backbone_obj=None, pbar=None):
+                                   backbone_obj=None, pbar=None, target_device=None, target_dtype=None):
         """Track with optional per-frame detection. Returns [N, max_N_obj, H, W] mask logits."""
-        N, device, dt = images.shape[0], images.device, images.dtype
+        if max_objects <= 0 or max_objects > self.INTERNAL_MAX_OBJECTS:
+            max_objects = self.INTERNAL_MAX_OBJECTS
+        N = images.shape[0]
+        device = target_device if target_device is not None else images.device
+        dt = target_dtype if target_dtype is not None else images.dtype
+        size = self.image_size
         output_dict = {"cond_frame_outputs": {}, "non_cond_frame_outputs": {}}
         all_masks = []
         idev = comfy.model_management.intermediate_device()
@@ -1656,7 +1673,7 @@ def track_video_with_detection(self, backbone_fn, images, initial_masks, detect_
                 prefetch = True
             except RuntimeError:
                 pass
-        cur_bb = self._compute_backbone_frame(backbone_fn, images[0:1], frame_idx=0)
+        cur_bb = self._compute_backbone_frame(backbone_fn, _prep_frame(images, slice(0, 1), device, dt, size), frame_idx=0)
 
         for frame_idx in tqdm(range(N), desc="tracking"):
             vision_feats, vision_pos, feat_sizes, high_res_prop, trunk_out = cur_bb
@@ -1666,7 +1683,7 @@ def track_video_with_detection(self, backbone_fn, images, initial_masks, detect_
                 backbone_stream.wait_stream(torch.cuda.current_stream(device))
                 with torch.cuda.stream(backbone_stream):
                     next_bb = self._compute_backbone_frame(
-                        backbone_fn, images[frame_idx + 1:frame_idx + 2], frame_idx=frame_idx + 1)
+                        backbone_fn, _prep_frame(images, slice(frame_idx + 1, frame_idx + 2), device, dt, size), frame_idx=frame_idx + 1)
 
             # Per-frame detection with NMS (skip if no detect_fn, or interval/max not met)
             det_masks = torch.empty(0, device=device)
@@ -1687,7 +1704,7 @@ def track_video_with_detection(self, backbone_fn, images, initial_masks, detect_
                 current_out = self._condition_with_masks(
                     initial_masks.to(device=device, dtype=dt), frame_idx, vision_feats, vision_pos,
                     feat_sizes, high_res_prop, output_dict, N, mux_state, backbone_obj,
-                    images[frame_idx:frame_idx + 1], trunk_out)
+                    _prep_frame(images, slice(frame_idx, frame_idx + 1), device, dt, size), trunk_out)
                 last_occluded = torch.full((mux_state.total_valid_entries,), -1, device=device, dtype=torch.long)
                 obj_scores = [1.0] * mux_state.total_valid_entries
                 if keep_alive is not None:
@@ -1702,7 +1719,7 @@ def track_video_with_detection(self, backbone_fn, images, initial_masks, detect_
                     current_out = self._condition_with_masks(
                         det_masks, frame_idx, vision_feats, vision_pos, feat_sizes, high_res_prop,
                         output_dict, N, mux_state, backbone_obj,
-                        images[frame_idx:frame_idx + 1], trunk_out, threshold=0.0)
+                        _prep_frame(images, slice(frame_idx, frame_idx + 1), device, dt, size), trunk_out, threshold=0.0)
                     last_occluded = torch.full((mux_state.total_valid_entries,), -1, device=device, dtype=torch.long)
                     obj_scores = det_scores[:mux_state.total_valid_entries].tolist()
                     if keep_alive is not None:
@@ -1718,7 +1735,7 @@ def track_video_with_detection(self, backbone_fn, images, initial_masks, detect_
                             torch.cuda.current_stream(device).wait_stream(backbone_stream)
                             cur_bb = next_bb
                         else:
-                            cur_bb = self._compute_backbone_frame(backbone_fn, images[frame_idx + 1:frame_idx + 2], frame_idx=frame_idx + 1)
+                            cur_bb = self._compute_backbone_frame(backbone_fn, _prep_frame(images, slice(frame_idx + 1, frame_idx + 2), device, dt, size), frame_idx=frame_idx + 1)
                     continue
             else:
                 N_obj = mux_state.total_valid_entries
@@ -1768,7 +1785,7 @@ def track_video_with_detection(self, backbone_fn, images, initial_masks, detect_
                     torch.cuda.current_stream(device).wait_stream(backbone_stream)
                     cur_bb = next_bb
                 else:
-                    cur_bb = self._compute_backbone_frame(backbone_fn, images[frame_idx + 1:frame_idx + 2], frame_idx=frame_idx + 1)
+                    cur_bb = self._compute_backbone_frame(backbone_fn, _prep_frame(images, slice(frame_idx + 1, frame_idx + 2), device, dt, size), frame_idx=frame_idx + 1)
 
         if not all_masks or all(m is None for m in all_masks):
             return {"packed_masks": None, "n_frames": N, "scores": []}

comfy_extras/nodes_ar_video.py

@@ -2,13 +2,15 @@
 ComfyUI nodes for autoregressive video generation (Causal Forcing, Self-Forcing, etc.).
   - EmptyARVideoLatent: create 5D [B, C, T, H, W] video latent tensors
   - SamplerARVideo: SAMPLER for the block-by-block autoregressive denoising loop
+  - ARVideoI2V: image-to-video conditioning for AR models (seeds KV cache with start image)
 """
 
 import torch
 from typing_extensions import override
 
 import comfy.model_management
 import comfy.samplers
+import comfy.utils
 from comfy_api.latest import ComfyExtension, io
 
 
@@ -71,12 +73,62 @@ def execute(cls, num_frame_per_block) -> io.NodeOutput:
         return io.NodeOutput(comfy.samplers.ksampler("ar_video", extra_options))
 
 
+class ARVideoI2V(io.ComfyNode):
+    """Image-to-video setup for AR video models (Causal Forcing, Self-Forcing).
+
+    VAE-encodes the start image and stores it in the model's transformer_options
+    so that sample_ar_video can seed the KV cache before denoising.
+    Uses the same T2V model checkpoint -- no separate I2V architecture needed.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="ARVideoI2V",
+            category="conditioning/video_models",
+            inputs=[
+                io.Model.Input("model"),
+                io.Vae.Input("vae"),
+                io.Image.Input("start_image"),
+                io.Int.Input("width", default=832, min=16, max=8192, step=16),
+                io.Int.Input("height", default=480, min=16, max=8192, step=16),
+                io.Int.Input("length", default=81, min=1, max=1024, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=64),
+            ],
+            outputs=[
+                io.Model.Output(display_name="MODEL"),
+                io.Latent.Output(display_name="LATENT"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model, vae, start_image, width, height, length, batch_size) -> io.NodeOutput:
+        start_image = comfy.utils.common_upscale(
+            start_image[:1].movedim(-1, 1), width, height, "bilinear", "center"
+        ).movedim(1, -1)
+
+        initial_latent = vae.encode(start_image[:, :, :, :3])
+
+        m = model.clone()
+        to = m.model_options.setdefault("transformer_options", {})
+        ar_cfg = to.setdefault("ar_config", {})
+        ar_cfg["initial_latent"] = initial_latent
+
+        lat_t = ((length - 1) // 4) + 1
+        latent = torch.zeros(
+            [batch_size, 16, lat_t, height // 8, width // 8],
+            device=comfy.model_management.intermediate_device(),
+        )
+        return io.NodeOutput(m, {"samples": latent})
+
+
 class ARVideoExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
         return [
             EmptyARVideoLatent,
             SamplerARVideo,
+            ARVideoI2V,
         ]