Memory between frames
Robust Video Matting keeps a hidden state across the clip, which is the structural fix for edge flicker rather than a smoothing filter applied afterwards.
Advertisement
Green-screen results without a green screen. The People model runs Robust Video Matting, a recurrent network that tracks the subject across frames, so edges stay stable instead of flickering. For products, pets, and everything else, two per-frame models cover any subject. Export a genuinely transparent WebM (real video alpha you can drop into any editor), a composited MP4 with a new solid, blurred, or image background, and keep the original audio. Every frame is processed on your device: the clip never leaves your browser.
Drop video files here
MP4, WebM, MOV, AVI · max 60s · processed at up to 1280px · never uploaded
Settings
Robust Video Matting, built for people and portraits. Tracks across frames, so edges stay rock-stable (no flicker). ~15 MB model, fastest of the three.
WebM: transparent VP8 alpha. Best for editing software.
Result proof
Input
Source clip + audio
Local pass
Tracked matte per frame
Result
Alpha WebM or composited MP4
How to verify it: Play the full result through hair and fast movement, place alpha output over both light and dark colours, and confirm the original audio remains in sync.
Temporal stability
Running a still-image matte on every frame independently is what makes cutout video shimmer: each frame guesses the boundary slightly differently, and the eye reads that as a crawling edge. The People path uses a recurrent model that carries state forward, so the boundary is a continuation of the last frame rather than a fresh guess.
yuva420p
True alpha WebM
Robust Video Matting keeps a hidden state across the clip, which is the structural fix for edge flicker rather than a smoothing filter applied afterwards.
Products, pets and objects fall back to per-frame segmentation with temporal smoothing, less stable than the recurrent path, but far steadier than raw per-frame output.
The WebM carries an actual alpha plane, so an editor reads the transparency directly instead of you keying out a baked-in background.
Know before you process: Sixty seconds at up to 1280px is the reliable bound, every frame runs inference and is then re-encoded on your device. Trim first and process the section you need.
Talking-head video without a green screen
Presenters, course modules and product explainers filmed against whatever wall was available. The People model is built for exactly this case, and it is also the fastest of the three, a recurrent network is doing less work per frame than a full segmentation pass, not more.
Overlays for social and streaming
A transparent WebM drops straight onto a scene in OBS or a timeline in an editor with no keying step. Because the alpha is real, semi-transparent edges composite correctly over a busy background instead of showing a hard fringe.
Product clips on a clean backdrop
Turntable and unboxing footage shot on a cluttered desk, recomposited onto a solid colour so a catalogue stays visually consistent. Use an any-subject tier here, Robust Video Matting is trained on people and will not reliably find a kettle.
Replacing a distracting background
Blur the original rather than removing it entirely when the room gives useful context but pulls focus. That path composites to MP4, which plays everywhere, and keeps the original audio track intact.
How it works
People, if your subject is a person (talking-head, presenter, dance clip), Robust Video Matting gives the most stable edges and is also the fastest. Pick an Any-subject tier for products, animals, or objects: Fast for speed, Best for the sharpest edges and hair.
Yes, it is encoded with an alpha plane (yuva420p), not a black or checkerboard background baked in. Chromium browsers and compatible video editors read the transparency directly. For broader playback compatibility, use a composited MP4.
Every single frame runs AI inference and is then re-encoded on your device. Sixty seconds at up to 1280px is the bound that finishes reliably without exhausting browser memory. Trim longer footage first (the video editor can do that) and process the section you need.
No. The models download to your browser once and run locally; frames, audio, and the finished file never touch a server. You can even process offline after the first model download.
Yes, choose a solid colour or a blurred version of the original under Background, and the tool composites it into an MP4. For full scene control, export the transparent WebM and layer it in the video editor.
Related tools
Free, private, no signup.
Remove a video background →