窗口锚定

The problem this is aimed at

Every long Wan video is glued together. The open base is frozen at Wan 2.2, SCAIL-2 is native at 81 frames, and past that everything gets chunked into sliding context windows with overlap and a fusion pass. Fusion is a weighted average, and averaging only works when both windows agree. When they don't - each has its own read on where the character is and how fast they're moving - you get the classic seam: soft motion, doubled limbs, two bodies in the overlap. It's the most-reported failure of chunking, and it clusters past the 81-frame mark.

Why they disagree: animate2 and SCAIL-2 have no skeleton, so nothing tells a window "the person was here". Each window's RoPE restarts at zero, and on SCAIL-2 driving frames are rigidly paired to output frames by position, so a window can only infer its baseline from its own slice of driving video. Two windows, two baselines, one average. This node works the content layer instead: it can't fix RoPE, but it can hand every window the same reference material, and stop two windows from both writing the same frames.

What it actually does

Wire the model output of Custom Context Windows (Manual) or Wan SCAIL-2 Context Windows into model, and its output onward to your KSampler. The node clones the model and wraps the handler's methods. With no context handler present it refuses outright rather than quietly doing nothing useful.

Three interventions, all optional.

Anchor frames (anchor_frames) - a few frames off the start of your driving video that every window shares as common context. Feed five, get five; there's no count parameter. They're VAE-encoded, so vae is mandatory when you connect them - the same Wan VAE as your conditioning node. With driving_only=true (the default) they're appended as extra driving tokens and repositioned by a RoPE patch to sit before the window: no frame shift, no bleed. Set it false and they're prepended into both the driving stream and the generation sequence, per-frame paired (closest to training semantics), at the cost of leakage since those are source-person frames; anchor_image_blank=true keeps their slots as noise placeholders instead. animate2 forces driving_only off regardless, because it hard-checks pose = generated frames − 1.

Read-only context (head_context_frames, latent frames) - each window may look at a few real neighbour frames: attention yes, fusion no. With overlap = 0 (schedule batched, or standard_even to flatten the tail) you get what the author calls one author per frame. The tooltip's own measurement: 2 is fine; 6 pushes per-window tokens to 27 against a training length of 21 and deforms limbs. Stay at 2–3. No tail version exists - it was removed, since those predictions get trimmed before fusion anyway.

Long-range memory (memory_frames, memory_stride, latent frames) - appends pre-window history driving frames as extra tokens, spaced per memory_stride (0, the default, spreads them across the whole history so a long clip still sees its own start). SCAIL-only; elsewhere it warns and switches itself off. Motion, not identity.

Also here: anchor_scope (fresh_only skips the window that already owns the pinned continuation frames), shared_head, and overlap_scope=single, which demotes already-written overlap to read-only. debug_log prints the divergence between adjacent windows in the overlap - a number for how ghosty this is - and flips without a restart, so it's the knob for A/B runs. Note that overlap_scope=single alone engages the node; with nothing connected and every count at zero, the model passes through untouched.

The honest read

The author labels this experimental and states plainly that its effectiveness and side effects aren't systematically verified. The changelog backs that up: an earlier version had the driving_only logic inverted, so SCAIL-2 users who believed they were on the no-frame-shift path were on the align path.

So treat it as an experiment with a protocol. Fix several seeds, and for each one render twice - node on, node off, one variable changed. Start with overlap_scope=single, one switch aimed straight at the averaging. Then anchors from the first 5–9 pixel frames of the driving clip (plus matching-frame anchor_mask_frames on SCAIL-2), then memory_frames from 4 to 8. Read the overlap divergence next to the picture, and don't trust a single run: the node's internal noise is a fixed seed, so results are deterministic per seed and config - "it worked on some seeds" is a real difference rather than luck.

Inputs, output, install

Required: model (from a context window node) plus the widgets above. Optional: anchor_frames (IMAGE), anchor_mask_frames (IMAGE, SCAIL-2 only), and vae. Output is a single model, straight into your KSampler. The errors you'll hit are all skipped steps: no handler on the model (node placed before the context window node), anchors without a VAE, or SCAIL-2 anchors without matching masks.

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

Or search ComfyUI-CustomNodeKit in ComfyUI Manager and restart. The pack's install.py pip-installs groundingdino-py and transformers on startup for its unrelated pose nodes - slow on first launch, not fatal if it fails. Last thing: the display name promises a two-sided "halo", but only the head side does anything.