72-Second Character Video Generation with ComfyUI Ref2VA VSA
Summary
- ComfyUI custom node that generates 5-second character videos in ~72 seconds on a single consumer GPU — vs the typical 10-minute standard
- Uses sparse attention applied to the MiniMax H3 video model to drastically cut compute
- Key trick: aggressively ignores most video data to save VRAM/compute, but keeps the reference character image fully intact so faces stay consistent
- Aimed at animators who want fast, consistent character clips without expensive cloud GPU rentals
- Very new repo (created Sep 5, 2026), rapidly gaining stars