Cutting out without a green screen: what AI can really do today — and what it costs
"With AI you don't need a green cloth any more." That sentence is in almost every brochure by now. It is true — partly. What never stands next to it are the numbers everything hinges on in the hall: how long a cut-out takes, on which hardware, and what happens when the two don't match.
We measured it, on four real camera photos at 30 megapixels. The results are less comfortable than the brochure.
What "AI cut-out" actually is
A model estimates for every pixel how much of it belongs to the subject — not just yes or no, but proportionally. That is exactly what matters with hair: a single strand covers thirty percent of a pixel, and if you turn that into a hard yes or no you get the cardboard cut-out everyone recognises immediately.
There are two families. Small models run in milliseconds and ship inside the software. Large models see considerably more at the hair edge, but weigh half a gigabyte and compute accordingly longer.
The number nobody quotes: processing time
Measured on a 30-megapixel photo, the model's pure processing time in each case:
- Built-in model: 0.2 seconds with a graphics card — 0.85 seconds without.
- Large model: 2.5 to 3.0 seconds with a graphics card — 23 seconds without.
Same file, same program code, a factor of eight apart. That is not decided by the software but by the machine under the table. And 23 seconds is not "a bit slower": that is how long a guest stands in front of the booth without knowing whether anything else is coming.
So the question to put to a supplier is not "Can you do AI cut-out?" but: "How long does it take on exactly the machine I am going to use?" Anyone who cannot give you a measured number has never checked. With BoothDock the time is shown next to every model — measured on the booth itself, not copied from our documentation.
More resolution makes it worse, not better
The obvious thought: if the edge is messy, compute at a higher resolution. The opposite is true — and we made this mistake ourselves.
A cut-out model is trained on a particular image size, typically 512 pixels. Give it 2048 and you don't change its acuity but its field of view: it only sees crops and no longer recognises the person as a whole. The result falls apart into fragments. A slider that goes up to 4096 is therefore not a quality feature but a trap.
The reverse doesn't hold either. At 512 the same model lost an entire shoulder on a 30-megapixel photo — dark jacket against dark foliage, too little difference for the model. In both cases 1024 was right. If someone shows you a quality slider, ask what the model was trained on. Everything above that is not headroom but risk.
The order of operations most systems get wrong
Almost every photo booth offers filters — black and white, sepia, punchy colours. And almost everywhere the filter is rendered into the picture first and the cut-out happens afterwards. That is exactly the wrong way round, for three reasons:
- Green-cast removal works with real green. After a warm-tone filter, the green is no longer green.
- Edge refinement uses the colour image to orient itself. Measured on real photos: with colour it improves the edge by fourteen percent, on a black-and-white image it makes it worse by twenty-one percent. A mono filter takes away exactly what it needs to work with.
- The model was trained on natural photographs. A "pop" look is image noise as far as it is concerned.
The right way: cut out first, then lay the look over the finished picture — then it also covers the new background, which looks more coherent anyway. At a demo, have a photo taken with a strong black-and-white filter and a cut-out. At the hair edge you will see immediately which order is being used.
When the green cloth still wins
AI does not replace the green screen everywhere. Three reasons still speak for the cloth:
- The guest sees themselves in the background beforehand. A chroma key runs live in the viewfinder, with practically no delay. A large AI model needs close to three seconds — too slow for a live preview.
- It runs on any hardware. A green cloth needs no graphics card, it needs even light.
- It is predictable. Well-lit green always gives the same result. AI is sometimes better and sometimes worse on difficult subjects.
AI wins where there is no room for a cloth, where the setup has to stand in five minutes — and with guests wearing hats or glasses or holding props up, which chroma key tends to fail on.
What to look for when buying
- Ask to see the measured processing time on your hardware, not on the supplier's demo machine.
- Ask whether there is a fallback to the fast model when the machine is too slow — otherwise the guest waits.
- Check the order of filter and cut-out on a picture with a strong look.
- Ask about the training resolution before you drag a quality slider up.
- Ask to see a result at full size, not as a thumbnail. At eight hundred pixels wide, every cut-out looks good.