Video background remover

MODNet is a portrait matting model: it was trained on people, so it is at its best on the clips creators film most. A talking head at a desk, a presenter pointing at the camera, a dance or workout filmed from a few steps away. Hair and shoulders come out with soft, natural edges rather than a hard cardboard outline, because the model gives every pixel a share of foreground instead of a yes or no.

It struggles where the person is not the clear subject. A product held out at arm’s length can be partly dropped, a pet on its own is treated as background, and someone tiny in a wide shot may vanish. Fast movement that smears into motion blur, and clothes that match the wall behind, also make the edge wander.

Flicker smoothing deals with the small frame-to-frame shimmer that every per-frame cut-out has. Each mask is blended with the one before it where the change is small, and passed straight through where the change is large, so a still shoulder stops buzzing while a waving hand doesn’t leave a ghost behind it.

Tip: Film against a plain wall with light on your face rather than behind you. Good contrast between you and the background does more for the edge than any setting here.

Browsers can’t record see-through video, so there is no transparent export. If you want to move the cut-out around, add text behind it or layer it over other footage, export on the green screen and use your editor’s chroma key effect. Once a clip is analysed you can switch backgrounds and export again: only the real-time recording runs a second time.

• Clips can be up to 20 seconds long. Every frame is analysed and its mask kept in memory until the recording is done; a 20-second clip at 30 frames a second is 600 masks.
• Anything above 1080p is scaled down to fit 1920 × 1080, in either orientation. Smaller videos keep their size.
• With WebGPU, the analysis runs at roughly 10 to 15 frames a second on a recent laptop, so a 10-second clip is ready in well under a minute. Without WebGPU every other frame is analysed and it takes several times longer.
• The recording plays the clip once in real time. Keep the tab in front and, on a phone, the screen on: browsers stop drawing video in background tabs, and the tool throws that attempt away rather than give you a frozen video.
• Chrome, Edge and Safari save an MP4. Firefox can only record WebM, which most editors open but you may need to convert before posting.
• Nothing is uploaded. The model downloads once from Hugging Face and your browser keeps it for next time.