SCAIL-2 in ComfyUI: setup and motion transfer tutorial

This tutorial covers SCAIL-2 in ComfyUI, with setup instructions and two character animation tests. The presenter describes it as an open source model that transfers motion directly from a driving video without a traditional pose extraction pipeline. The supplied workflow still uses segmentation masks to select characters.

Setup covers FP16 and FP8 weights, plus GGUF options for limited VRAM. Required components include a UMT5-XXL text encoder, CLIP Vision H, the Wan 2.1 VAE and a SAM 3.1 multiplex checkpoint. The presenter says the distilled LoRA supports generation in as few as six sampling steps. Supported output resolutions are 512p and 704p, with 704p recommended for pose-driven and replacement workflows.

Replacement mode uses a reference character within the driving video's scene. Standard animation mode takes motion from the video and character appearance and scene information from the reference image. Controls include character count, frame limits and a sampling window for longer clips. The presenter recommends matching the reference image closely to the driving video's first frame.

The tests replace a man in a kitchen dance scene and transfer two boxers' movements to martial artists. The presenter reports preserved backgrounds and character interactions in these examples; the results do not establish reliability across other inputs.