VACE

Why we included this project

Video teams that need both generation and editing often end up stitching several models together. VACE folds that into one codebase: reference-to-video, video-to-video, and masked editing all run through the same inference path, so you can swap an object, move a subject, expand a frame, or animate a still without switching tools. The model builds on Wan2.1 and ships in 1.3B and 14B sizes, which lets you trade speed for quality depending on your hardware, and the same weights are published on HuggingFace and ModelScope. Its real strength is composition, since you can chain several operations in a single pass instead of round-tripping through separate generators. The repo includes official inference code, a user guide, and the VACE-Benchmark evaluation data, which makes reproducing the paper's results fairly straightforward.