Wan SCAIL Context Windows

The generic context-window node (WanContextWindowsManual_Custom in this same pack) handles plain conditioning, but SCAIL-2's conditioning is a mess of differently-shaped fields: reference latents that must stay global, a pose-video latent, and two multi-channel masks (ref_mask_28ch, driving_mask_28ch) whose time axes live on different dimensions. Slice those naively and you'll break the model. WanSCAILContextWindows is the SCAIL-specific wrapper that knows exactly which field lives where and slices each one correctly per window.

It's the equivalent of what the community's single-node chunkers (SCAIL-2 Infinity, scail-auto-extend) do - hiding the 81-frame windowing math behind one node - but tuned to this pack's conditioning format. You wrap your SCAIL-2 model, feed the wrapped model to the KSampler alongside the conditioning from WanSCAILToVideoMultiRef, and windows just happen.

The inputs that matter

• model - the SCAIL-2 model to wrap (required).
• context_length (default 81) - window length in pixel frames; the tooltip recommends matching your SCAIL length parameter exactly (81). It's converted to latent frames internally.
• context_overlap (default 30) - overlap between windows, pixel frames.
• context_schedule - static_standard is the recommended option here (fixed windows); the uniform and looped schedules exist for other uses.
• context_stride / closed_loop - only affect the uniform/looped schedules.
• fuse_method - pyramid (default) is the recommended blend; relative, flat, overlap-linear also available.
• freenoise - FreeNoise noise shuffling for smoother inter-window blending.

The single output is the wrapped model.

How it's different from the manual node

Three deliberate differences, straight from the README: it automatically handles SCAIL-2's special conditioning fields - ref_mask_28ch keeps its reference part (the first N frames) and slices only the video portion, driving_mask_28ch slices entirely, and pose_video_latent slices on its time dimension. Second, there's no prefix-reference mechanism: SCAIL-2's reference latents and the first N frames of ref_mask_28ch are global by design, so the node keeps them intact across all windows instead of re-prepending them. Third, there's no causal_window_fix - SCAIL-2 does inter-chunk anchoring through the previous_frames mechanism instead, so serial-window continuity is the two-phase sampler's job, not this node's.

Pitfalls

Keep context_length in sync with the length you give the conditioning node - the whole design assumes windows match chunks. If you're feeding previous_frames for anchoring, that happens on the sampler/conditioning side; this node won't add it. And remember the cost reality of SCAIL-2: even with windows, 20 seconds of output is tens of minutes of compute on consumer hardware.

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. Works with your existing SCAIL-2 weights; nothing extra to download.