VACE

VACE (Video All-in-one Creation and Editing) is Alibaba Tongyi Lab's unified model for video synthesis, introduced in the arXiv paper of March 11, 2025 and open-sourced in May 2025 as Wan2.1-VACE. It organizes editing, reference, and masking inputs into a unified Video Condition Unit and injects task concepts through a Context Adapter, so one model handles reference-to-video generation, video-to-video editing, and masked editing with performance reported on par with task-specific models.

Alibaba released it in 14B and 1.3B parameter versions on Hugging Face, GitHub, and ModelScope, with the smaller version able to run on consumer-grade GPUs. Practical capabilities include pose transfer, motion and depth control, recolourisation, selective-area add/modify/delete, and spatio-temporal extension of video boundaries. Check the model card for the exact license before commercial use.