Green screen vs AI matting: which should you use?

Chroma keying is colour-based: the software makes every pixel close to a chosen colour transparent. Green is used because it's far from human skin tones and because camera sensors capture the most detail in the green channel. That simplicity is the strength, and the source of all the classic problems:

• Lighting is everything. The screen must be lit evenly; wrinkles, shadows and hot spots become different shades of green that the keyer can't remove cleanly. In practice you need two lights on the screen and separate lights on yourself.
• Spill. Light bounces off the backdrop and tints your hair, shoulders and skin green. Keyers have spill-suppression controls, but heavy spill (standing too close to the screen) never keys perfectly.
• You can't wear green. Green clothing, green props, sometimes even green-ish eyes get keyed into holes.
• Space and cost. A usable setup is a backdrop (fabric or paper), a frame or wall mount, and at minimum two softbox lights: realistically €80–300 and a couple of square metres you have to keep clear. Setup and teardown take time every session.

In exchange, you get something valuable: a key that works on anything in front of the screen (people, pets, products, smoke, glassware) and that runs in real time, which is why broadcast weather studios and streamers with a permanent setup still use it.

AI matting doesn't look at colour: it looks at content. A neural network trained on huge amounts of footage of people predicts, for every pixel of every frame, how much of it belongs to the person (a soft alpha value between 0 and 1, not a hard yes/no). The model behind VideoBGNinja, Robust Video Matting (RVM), adds a crucial ingredient for video: temporal tracking. It carries a recurrent memory from frame to frame, so it remembers where you were a moment ago. That's what keeps edges stable while you move, instead of flickering like per-frame photo cutouts do.

The practical consequences:

• Zero setup. Film in your living room, office or outdoors. No backdrop, no lighting rig, no spill, and you can wear green if you like.
• Genuinely good hair edges. Because the model outputs soft alpha, fine hair strands get partial transparency, comparable to a decent chroma key, without the green fringe.
• Zero marginal cost. No hardware to buy. Tools like VideoBGNinja run the model on a GPU server and give you a transparent WebM, a green-screen MP4 and an alpha matte for free.

Most "AI matting is worse" complaints come down to a fixable filming problem, not the model. The common ones:

• Edges shimmer frame to frame. Usually a handheld camera — the temporal tracker locks on best against a static background. Put the phone on anything stable and the wobble largely vanishes.
• Hair goes soft or grey. Low contrast between hair and the wall behind it, or too little light. Add a lamp and step away from the wall; this is the one area a perfectly-lit green screen can still edge out.
• Chunks of you drop out. Dark clothing against a dark room reads as background, so raise the light or add some tonal contrast until the two clearly differ.
• The result is nearly empty. There's no clear person in frame, or the subject is an object or pet, which sits past what a human-matting model can do. For those, a green screen remains the right tool.

A deeper walkthrough of getting it right first time is in how to film yourself for a clean cutout.

The same "is it the model or is it fixable" question comes up with old-photo restoration — PicReviver's honest guide covers what AI upscaling, colourising and face restoration actually do, and where the limits are.