Remove a video background, free

What exactly does this tool do?

It walks through every single frame of the video and finds, on each one, where the foreground is — usually a person — and where the background is. That decision is made by a small neural network, the same one behind our image background remover. It then replaces the background with whatever you chose: a flat colour (green screen, blue screen, white, or any colour of your own), or a picture you uploaded.

None of this needs a green screen, lights or a studio: the whole point is that you can lift yourself out of an ordinary phone recording shot against an everyday background. The finished video is an MP4, so it can be uploaded anywhere, and we do not touch the audio — it is carried over unchanged, with no re-encode. We do not produce a transparent (alpha-channel) output yet: if you are taking it into an editor, choose the studio green and key it out there — which is exactly why we use the real keying colour.

What it can and cannot do

Let us be honest: this is not studio-grade keying, and it was never meant to be. The model works at 320×320 pixels, so the fine detail of the outline — individual hairs, glasses frames, the gap between fingers, translucent fabrics — blurs, or stays attached to the background. What works well: one person, roughly facing the camera, standing out from the background. What is hard: several people in one shot, a moving background, fast motion, dark footage, and any object the model does not usually see as a foreground.

Video has a second failure that only shows up here: because every frame is a separate decision, the boundary sits slightly differently on two consecutive frames, and the outline „boils”. We damp that with temporal smoothing — the mask is blurred and averaged with the previous frame's, while at a cut the average is thrown away so no ghost is left in the new scene. It does not remove the effect entirely, but it makes the edge far calmer.

In v1 the clip can be at most 30 seconds and the output at most 720p. Both are cost limits: every frame is a separate run, and a 30-second clip is 900 of them. With a graphics card that is typically about a minute; without one it can be over a quarter of an hour — which is why we ask for a separate confirmation before starting on the processor.

How do you get a good result?

• Stand away from the background, and make sure it is lighter or darker than your clothes — the model works from contrast.
• Even light beats backlight: if the sun is behind you, you become a silhouette and the outline blurs too.
• A fixed camera helps. A hand-held phone's moving background is different on every frame, and that makes the mask restless too.
• If you are taking it into an editor, choose the studio green or the studio blue: those are the real keying colours, so the editor's default setting keys them out first time.

Frequently asked questions

Is my video uploaded anywhere?

No. The clip stays in your browser's memory and the model runs on your own machine. Nothing reaches FileMv unless you ask for the shareable link at the end yourself.

Why does it need a graphics card?

Because it is 900 separate runs. With WebGPU each frame is milliseconds and the whole thing takes seconds; without it the processor does the same work, and a 30-second clip can go past a quarter of an hour. We do let you start it, but only after you have seen that number.

Can I get a video with a transparent background?

Not yet. MP4 effectively cannot carry an alpha channel; WebM with VP9 can — but the browsers' built-in encoder cannot encode an alpha channel, so it would have to be done in software, which our own measurements put orders of magnitude too slow to wait for here. Until that changes, an MP4 on studio green is the honest answer: any editor can key that out.

Why only 30 seconds?

Because the processing happens frame by frame, and a longer clip would load your machine for minutes. If you have longer material, cut the important part out with the video trimmer and bring it back — trimming is done in seconds, with no re-encode.

Can it cut out several people at once?

Sometimes, but it is not built for it. The model looks for a single salient foreground, so out of two or three people it usually finds the largest and nearest one reliably, and the others may partly stay in the background.