WAN Context Windows (Manual Custom)

WAN-class video models have a native context of about 81 frames - roughly 5 seconds at 16fps. Everything longer has to be chunked: split the generation into overlapping windows, denoise each, and stitch. That arithmetic (81-frame windows, overlap, 4n+1 frame counts) is exactly the manual tax the community spent 2026 trying to remove. This node is this pack's answer for WanAnimate: a model wrapper that does the windowing for you, with two extras the stock chunkers don't have.

It wraps your MODEL and hands back a patched MODEL for the KSampler. Inside, it slices the latent's time dimension into overlapping windows according to a schedule, denoises each, and fuses the results with a weighting method. All context_length and context_overlap values are in pixel frames - the node converts to latent frames internally (every 4 pixel frames ≈ 1 latent frame), which is the single most common source of confusion.

The inputs that matter

• context_length (default 81) and context_overlap (default 30) - window size and overlap, pixel frames.
• context_schedule - how windows are placed: standard_uniform (evenly spaced, no repeats) is the default pick for most work; looped_uniform iterates; standard_static is a fixed stride; batched is no-overlap quick batch processing.
• fuse_method - how overlapping regions blend: pyramid (center-weighted, the default and usually the best), flat, overlap-linear, or relative.
• freenoise - FreeNoise-style noise shuffling across windows; helps texture continuity when windows overlap little. Costs a bit of time.
• prefix_latent_num - the number of reference frames to prepend to every window. Set it to the count of your reference images; each encodes to one latent.
• causal_window_fix (default on) - the interesting one. In serial mode it prepends each window with the previous window's denoised last frame, so interaction details like footprints on the ground survive across windows. The cost is parallelism: windows run serially inside each denoising step.

The dynamic prefix (the part nothing else does)

Besides the static prefix above, this node can swap reference images per window based on yaw. Connect window_reference_images (from Reference Image Selector's raw_reference_images), latent_yaw_angles (from WanAnimateToVideoCustom's latent_yaw_angles), vae, and reference_angle_map - then each window picks the reference angle that best matches that segment's facing direction, so a turn-around gets the back-view reference instead of the front one. allow_switch_main governs whether the main reference can be swapped; background_images feeds the 1+4n tiling.

Common pitfalls

Everything here is in pixel frames except prefix_latent_num, which is in latent frames - don't multiply by 4 there. causal_window_fix on means serial execution; if you're wondering why a long generation is slow, this is why, and it's usually worth it. And the dynamic prefix silently doesn't activate unless you've connected all four inputs - the console prints a note when it's off, so check the log if your reference images aren't rotating.

Install

Part of ComfyUI-CustomNodeKit. ComfyUI Manager → search "ComfyUI-CustomNodeKit", or:

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

Restart ComfyUI. No model files bundled.