Context Window Anchor (窗口首帧锚定)

The problem it's aimed at

Wan's native context is 81 frames, so anything longer is windowing: slice the sequence into overlapping windows, sample each against the same model, fuse the results with a weight schedule. That works, but it has a specific failure. Animate2 and SCAIL-2 derive their "reference image ↔ driving video" body-scale and position alignment per window, from whatever driving frames that window contains. Window one starts on one pose, window four on another, so each window anchors itself differently - and the fusion shows it as a jump at the seam.

Long-context runs generate more error than answer in general - window-size mismatches that still error out, and Expected all tensors to be on the same device blow-ups the moment you plug reference latents into a context setup (1pao1t6, 1pc8b7z). And the author labels this node experimental and test-only: effectiveness unproven, side effects on quality, identity and sampling stability unevaluated.

How it works

It edits conditioning, at the point where the context handler re-slices each window - no pixels, no latents. The required model input has to be the output of a context-window node (Custom Context Windows (Manual), or Wan SCAIL-2 Context Windows for SCAIL-2), because those are what stash a handler on model.model_options["context_handler"]. This node clones the model, shallow-copies that handler and wraps its get_resized_cond, so after the per-window slice, windows after the first get their leading pose_latents frames overwritten with window 0's (and on SCAIL-2, the driving mask sam_latents too).

Three consequences fall out of doing it there rather than on the latent:

• Every sampling step, because get_resized_cond runs every step.
• Never changes tensor lengths, so animate2's pose_latents.shape[2] == f_gen - 1 check still passes.
• Never crosses chunks: the anchor source is window 0 of the current segment, so drift between segments is untouched. It's a within-segment seam fix, and the author says so.

The anchor_offset default of -1 is the clever bit. Context nodes prepend the prefix reference frames to each window's index list, and the driving-video slice comes from that same list - so every window's slice starts with P-1 prefix-derived frames that are identical across windows, and anchoring there would do nothing. Auto offset skips them: six reference images → offset 5.

The inputs worth setting

• anchor_frames (default 1, max 32, 0 disables) - start at 1, try 2–3 after that. Turn on log_debug for one run: it prints every window hit with shapes and first-frame values.
• anchor_targets - "驱动视频 + 驱动遮罩 (auto)", "pose only" or "mask only". On SCAIL-2, cycling through these shows which of the two is dragging your seam around.
• anchor_steps vs blend - two strategies, not two ends of one slider. blend below 1.0 mixes the anchor with the window's own frames, feeding a synthetic in-between every step. anchor_steps = N anchors for the first N steps, then returns to real conditioning. Try blend=1.0 + anchor_steps=1 against blend=0.7 + anchor_steps=-1.
• global_pose_frames (plus global_mask_frames and vae on SCAIL-2) - hybrid anchoring. Feed it the head of the whole driving video, taken before your loop trims the clip to the current segment. More frames than the anchor band (M = ((T-1)//4)+1) means full replacement, fewer means partial. Optional - unconnected, it behaves exactly like plain per-segment anchoring.

Output is a single model: drop it between the context node and the KSampler.

Installing it

Same pack, same three steps:

cd ComfyUI/custom_nodes
git clone https://github.com/user2318/ComfyUI-CustomNodeKit.git
cd ComfyUI-CustomNodeKit
pip install -r requirements.txt

Restart, or search "ComfyUI-CustomNodeKit" in ComfyUI Manager. No model files, no extra dependencies - it's pure Python against the model object. (The pack's install.py does pull groundingdino-py and transformers for its unrelated detection nodes; if that fails, the rest loads anyway.)

Where people get burned

The node does nothing and doesn't complain. If the model doesn't carry a context_handler, it logs a warning and passes the model through - no error, no red node, and you conclude the setting didn't help. Usual cause: it's wired before the context node instead of after, or a Fast Muter upstream cut that node out of the path.

No change with log_debug on. Only tensors split per window get touched; whole-sequence conditions are skipped deliberately, and SCAIL's reference-frame tensors are in that category. Also check enabled, anchor_frames > 0 and anchor_steps != 0 - all three pass through when off.

blend won't soften the mask. On mask keys the blend is coerced to 1.0 by design: sam_latents is category data (colour coverage per identity), and mixing two frames of it produces boundaries that match neither. If the mask anchor causes jitter, use anchor_steps to switch it off early instead. A shape-mismatch warning means source and target shapes disagree and that key was skipped.

And the caveat that decides whether to use it at all: same seed, one variable at a time, node in versus node out, twice. If your seams already look clean, leave it out of the graph.