MatAnyone 2: ditch green screens with a 140MB video matting model

MatAnyone 2 is a lightweight open-source video matting model built to pull people out of footage cleanly without green screens or studio lighting. The project has been accepted to CVPR 2026, with code already out on GitHub.

What it does and how to use it

The model packs impressive matting precision into a tiny 140MB footprint. Even during fast camera motion or tricky shots with fine hair strands, it extracts clean Alpha mattes. Compared to existing baselines like GVM, edge sharpness—especially around hair—is noticeably better.

Testing it out takes zero setup. You can jump into the free interactive demo hosted on Hugging Face: just upload a clip, click on the person once, and the model automatically tracks and segments them across the entire video.

Why it matters

For indie creators and video editors, ditching green screens removes a massive headache around lighting, setups, and tedious rotoscoping.

At roughly 140MB, this model is lightweight enough to eventually fit straight into local desktop editors or mobile pipelines. If your workflow involves isolating moving subjects, this is an easy tool to test right now.