Why your background edges look bad, and how lighting fixes it

A segmentation model takes your camera frame and returns a map of how likely each part of the picture is to be a person. That map is much smaller than the picture — in this app it is 256 by 144 pixels standing in for a frame up to 1920 by 1080. A single mask pixel therefore covers a block of your image, and the boundary it describes is the model's general idea of your outline rather than your outline.

Stretch that mask up naively and you get a boundary several pixels thick that sits somewhere near where you are — which in practice means a rim of your real room surviving around you. That rim, against a replacement backdrop, is what people call a halo.

So the mask is not stretched naively. It is re-sampled using your camera frame as a guide: for each output pixel, nearby mask samples are weighted by how closely their brightness in the camera image matches the pixel being filled in. Samples from across a real edge in the picture contribute almost nothing, so the boundary is pulled onto the edge that is genuinely present. The technique dates back to work published for Google Meet and it is why modern background replacement looks far better than it did five years ago.

And there is the catch. The whole method depends on there being an edge in your picture to snap to. Where the boundary between you and your room is invisible — a dark room, a wall the tone of your hair, your face in shadow against a bright window — the guide has nothing to say, and the result falls back towards the coarse mask you were trying to improve on.

A flickering halo around your whole outline — not enough contrast between you and the room. Add light in front of you.

A shoulder or an arm briefly vanishing — the model itself is losing confidence, usually because you are dark or partly out of frame. Add light, and check you are not sitting at the edge of the picture.

A boundary that looks stepped or blocky — the mask's own resolution showing through, which happens where the guide finds a very strong edge and little else. Rarer, and generally improved by the same things.

Everything soft and slightly wrong all the time — you may be on a browser without WebGL2, in which case the app is compositing correctly but cannot run the refinement pass at all. Chrome, Edge, Firefox and Safari 16.4 or later all have it.

A stuttering, laggy preview — not an edge problem. That is your machine running out of frame budget; close other tabs and drop to 720p.

Use a blur instead of a replacement backdrop. Its errors are hidden by its own effect, and it keeps the real light of your room, which is the thing that makes you look photographed rather than pasted on. For anything where you cannot control the room — an interview, most obviously — that is the choice that fails most gracefully.

And if a perfect edge is a genuine requirement rather than a preference, the honest answer is a physically plain, evenly lit backdrop and a keying pass in an editor. Live segmentation is very good; it is not a substitute for controlling the shot.