VACE

One model covering reference-to-video generation, video-to-video editing, and masked video editing is the pitch of VACE, the Alibaba Tongyi vision lab framework accepted to ICCV 2025. Its unifying idea is the Video Condition Unit, an input format that packs text, reference images, source video, and spatiotemporal masks into a single conditional interface, so tasks such as pose transfer, outpainting, object swap, inpainting, and move-anything composition all run through the same weights instead of separate specialist models. Released checkpoints build on Wan2.1 at 1.3B and 14B parameters plus an LTX-Video 0.9 variant, with the 1.3B model aimed at consumer GPUs. Installation is a pip requirements setup on PyTorch 2.5.1 with CUDA 12.4, and the models reach the broader ecosystem through native Diffusers support and Kijai's ComfyUI wrapper, with VACE capabilities folded into the Wan model line. The Apache-2.0 repository, at about 3,900 GitHub stars, also publishes VACE-Benchmark evaluation data, making it the reference open toolkit for controllable video editing.