Removing a video background, with or without a green screen

Cutting a subject out of a video used to require a green screen and a chroma key. Modern background removal runs a segmentation model on every frame instead, which means it works on ordinary footage — provided you understand where it breaks down and shoot around those cases.

What automatic removal actually does

A segmentation model looks at each frame and decides, pixel by pixel, what is subject and what is background. It is not detecting colour like a chroma key; it is recognising shapes it was trained on, which for most models means people first and everything else second. This is why it handles a person against a cluttered room easily and struggles with a chair, a pet at an unusual angle, or two overlapping people. When it works it is astonishing, and when it fails it fails at the edges — hair, fingers, and anything semi-transparent.

Shooting so the model has an easy job

Three things make the difference. Contrast between subject and background: not colour contrast specifically, but tonal separation, so a dark jacket against a dark wall is the worst case. Even lighting on the subject, since deep shadow reads as background to most models. And a stable camera, because a shaky frame makes the mask jitter at the edges even when each individual frame is cut correctly. Loose hair and fine detail will always be the weak point; tie hair back if the cutout matters.

When to use a green screen anyway

Chroma key still wins for two situations: fine detail like frizzy hair or a glass, and long takes where a per-frame model would drift. A green screen needs to be evenly lit and far enough behind the subject to avoid green spill on shoulders and hair, which is the mistake that makes amateur keys look amateur. If you shoot green regularly, light the screen separately from the subject.

Fixing an imperfect cutout

You rarely need a perfect mask, because most meme videos composite the subject onto a busy background where edges are hard to inspect. If the edge does show, three cheap fixes work: feather the mask by a pixel or two so the transition is not razor sharp, add a subtle shadow under the subject so the eye reads it as sitting in the scene, and match the colour temperature between subject and new background. That last one fixes more perceived cutouts than any amount of mask refinement.