comfyui-scail-2-loop-sampler

ComfyUI SCAIL-2 Loop Sampler

SCAIL-2 Loop Sampler replaces the manually duplicated Base/Extend subgraphs in video_wan21_scail2_character_replacement_int8.json.

The reusable setup stays outside the loop: load the diffusion model and VAE, encode prompts and CLIP Vision, and create the full driving/reference colored masks. Connect those results to the loop sampler with a sampler and sigmas. The node reuses reference conditioning for full-size segments. A shorter final segment rebuilds it because the SCAIL-2 reference mask has a segment-length-dependent temporal shape. The node then repeats only the segment-dependent work:

1. Build native WanSCAILToVideo conditioning for the current frame offset.
2. Sample and decode that segment independently.
3. Drop the five overlapping anchor frames from extensions.
4. Apply the workflow's Reinhard seam color transfer.
5. Append the new frames and crop the result to the driving-video length.

The implementation uses ComfyUI's V3 node API (io.ComfyNode, define_schema, ComfyExtension, and comfy_entrypoint). It requires a ComfyUI build that contains the native WanSCAILToVideo and SCAIL2ColoredMask nodes.

For the supplied workflow, keep the original values of 81 for segment_length and 5 for previous_frame_count. Feed the images output of GetVideoComponents into pose_video; audio and FPS remain on the existing CreateVideo path.