Remove Video Background

Replace the background behind a person with a colour, an image or a blur, or remove it for a transparent WebM, in your browser without uploading the video.

Choose a video

or drop it here

The video is processed on this device and never uploaded.

MP4, MOV, WebM, MKV, AVI, WMV, FLV, 3GP, TS / MTS / M2TS or VOB / MPG

Files over 1 GB can be slow. Very large outputs need a browser that can save straight to a file.

How it works

Choose a video with a person in it and press Find the person. Your browser goes through every frame and works out which pixels belong to the person (a matte). This runs on your device, and the frames are never uploaded. Then pick the new background, check it in the preview and press Replace background. The browser draws every frame again over the new background and encodes a new file.

The person is found by MODNet, a portrait-matting model that runs on your graphics card (WebGPU) when it can and on the processor (WebAssembly) otherwise. On the first frames its result is compared with Google's MediaPipe selfie segmenter, and if the two disagree, MediaPipe is used for the whole video. The outline is smoothed over neighbouring frames so the edge does not flicker.

Supported files

Inputs: Videos your browser can decode with WebCodecs: MP4 and MOV (H.264, HEVC on most devices), WebM (VP8, VP9, AV1). AVI and FLV files with H.264 video are repacked with FFmpeg first (copied, not re-encoded); WMV and VOB are not supported here yet.
Outputs: MP4 (H.264 and AAC) or WebM (VP9 and Opus) at the original resolution; a transparent background is always WebM (VP9 with alpha).

Limitations

It works on people only. The models are trained on people, so products, pets and other objects are treated as background.
Edges are soft. Hair, fast movement, motion blur and busy backgrounds can leave halos or missing pieces, so check the preview.
Transparent output is WebM (VP9 with alpha). Safari support for alpha WebM is limited; it may show a black background.
Long videos take time and memory. The analysis keeps an outline for every frame (about 66 MB per minute at 30 fps). Above about 8 minutes, only every 2nd or 3rd frame is analysed and the frames between are blended.
The video is always re-encoded. Cropping, rotating or resizing cannot be combined with background replacement here.
The first run downloads the models (about 33 to 66 MB). Later runs use the browser cache and also work offline.

Questions

Is my video uploaded?

No. The analysis and the conversion both run in your browser on your device. The only downloads are the app and the models, from this site.

Why was no person found?

The models only recognise people. A video of a product, an animal or a landscape has no person in it, so everything counts as background. Very small or heavily backlit people can also be missed.

Which models are used?

MODNet (Apache-2.0) finds the outline, and the MediaPipe selfie segmenter (Apache-2.0) checks it and takes over when the two disagree. MODNet runs locally with ONNX Runtime Web and MediaPipe with MediaPipe Tasks.

What happens to the sound?

The sound is kept. Only the picture is redrawn, and every frame is re-encoded as MP4 (H.264) or WebM (VP9) at the original resolution.

Will the transparent video work in my editor?

Editors and browsers that read VP9 alpha in WebM (Chrome, Edge, Firefox, many video editors) show it without a background. Safari and some players show black instead.

Example

A 20-second webcam clip was recorded in a messy room. Press Find the person, choose Blur at 60 % and press Replace background. The output is the same clip with the room softly blurred behind you, like a video-call background, and the sound kept.