UniAVC: an open model that edits a video and its soundtrack from one instruction
Researchers at HKUST and Huawei released UniAVC, a 19B open model built on LTX-2.3 that generates and edits video and audio together, so removing a person also removes their sound. It needs about 39 GB of VRAM even with offloading, which rules out consumer GPUs for now.
What it does
One instruction can restyle or re-dress a subject, add or remove a person, move the scene to a new background, or change what someone says with matching lip movement. It also does text-to-video with sound, video for an existing audio track, and a soundtrack for a silent clip. An intent-aware mode locks one track (say the picture) while editing the other, and a post-refine pass re-runs only the half you want to improve.
Results and limits
The authors show side-by-side clips against InstructAV2AV and MiniMax H3 and say results are comparable to specialized models; there is no independent test yet. Defaults are 1280×704, 121 frames at 24 fps (about 5 seconds) and 30 steps, tested on one 80 GB GPU.
Worth knowing
Weights plus the Gemma-3 12B text encoder are about 65 GB, and both are gated on Hugging Face. UniAVC's own code is CC BY-NC 4.0 and the weights fall under the LTX-2 Community License. End your edit prompts with what to keep, as every official example does.
Try it yourself
Replace the background with [new location]. Keep the [person] and their speech unchanged.