How to blur or replace a video background

Five steps. Judge the result at your shoulders and hair — that is where segmentation fails first.

1. 1

   Pick a live camera or an uploaded file

   Live camera is the path for recording something new — a talking-head clip, a course segment, a call you want to rehearse. Upload is for a file you already have. Both run through the same segmentation and both record the same way, so the choice is only about where the frames come from.

2. 2

   Let the segmentation model load once

   The first run downloads a segmentation model, then keeps it cached for the rest of the session. Your video never leaves the machine — the model comes to the video rather than the other way round — but that first load needs a connection and a few seconds before the preview starts responding.

3. 3

   Choose blur, a colour, an image, or a video background

   Blur keeps your real room but throws it out of focus, which reads as professional and never looks like a bad composite. Colour gives you a flat backdrop. Image and video replace the background outright. Blur is the most forgiving of imperfect lighting because the edge between you and a blurred room is much less obvious than the edge between you and a sharp photograph.

4. 4

   Set the intensity while watching the edge, not the middle

   Judge the result at your shoulders, hair and any gap between your arm and your body — that is where segmentation fails first. If the outline flickers, more light on your face and more contrast against the wall behind you will do more than any slider.

5. 5

   Record, then download

   Recording captures the processed canvas as it plays, so it runs in real time: a three-minute clip takes three minutes to record. Stop when you are done and download immediately. Nothing is queued, nothing is rendered afterwards, and there is no watermark on the output.

How it works, and what that costs you

Everything staying on your device buys privacy and costs you speed and container choice. Both trades are worth knowing before you start a long recording.

A segmentation model finds the person, frame by frameMediaPipe · selfie segmenter

Each frame is passed through a model that outputs a mask separating a person from everything behind them. The tool then draws you from the original frame and your chosen background everywhere the mask says "not person". This is why it is reliable on a person facing the camera and unreliable on anything else: the model was trained to find people, not arbitrary objects.

Everything runs on your machine

Frames go from your camera or your file into a canvas in the page and never to a server. There is no upload step, no queue, and no copy of your footage anywhere but your own device. The one thing fetched over the network is the model file itself, which is a fixed asset and carries none of your content.

Capture is real time, not rendered

The preview is a canvas redrawn every animation frame, and recording captures that canvas live. There is no faster offline render path: a five-minute source video takes five minutes to record, and closing the tab mid-recording loses it. Plan around that for anything long, and record in sections if you can.

The download is WebM

Output is a WebM file, VP9 where the browser supports it. That plays natively in Chrome, Edge and Firefox and imports into DaVinci Resolve, Premiere and most editors. It is not MP4: if you need MP4 for a platform that insists on it, or for older software, convert the WebM afterwards with HandBrake or ffmpeg. No watermark is added at any point.