← All articles

BoothDock · Wiki

Cutting out – green screen and AI

BoothDock can replace the background in two ways: the classic route with a green cloth (chroma key), or with AI and no cloth at all. The template only decides whether the background is replaced and what it is replaced with. How good the result looks depends on the booth – which is why you tune it in the admin area under Camera, not in the design.

Which route suits you

Green cloth (chroma)AI cut-out
Guest sees themselves in the scene beforehandyes, in the viewfinderyes, for photos
Setupcloth plus even lightingnothing
Hardwaredoesn't mattergraphics card recommended
Hats, glasses, propsoften difficultusually fine

In short: if you have the space and the light, the cloth is the calmer choice. Where the setup has to stand in five minutes, the AI wins.

AI only cuts out photos; GIF, Boomerang and video stay as they are. The green cloth also replaces the background in GIF and Boomerang, but not in video.

Tuning on the test photo

Under Camera → Cut-out (section Tune cutout) you take a test photo and see every change on it immediately – not on your first guest.

  • Pipette: clicking into the green of the test photo picks up the key colour. It averages a 5×5 area so that sensor noise doesn't hand you a random neighbouring colour.
  • Tolerance decides how far a colour may differ from the key and still be removed.
  • Edge softness is the width of the transition. Too hard looks cut out with scissors, too soft frays.
  • Suppress spill with Spill strength (%) takes away the green shimmer the cloth throws onto hair and shoulders; Spill across whole subject works across the whole subject instead of only at the edge. This is the setting with the most visible effect – in the AI route it is called Suppress spill for AI too.
  • Show matte shows the bare mask instead of the picture. Mistakes show up there earlier than in the finished photo.

Choosing a model

For the AI cut-out, each model shows how long it took on this booth per photo – measured on the machine itself, not copied from a manual. Two measured values on a 30-megapixel photo:

  • Fast, good quality (built in, always available): around 0.2 s with a graphics card, 0.85 s without.
  • People & hair (slow) (downloaded, about 490 MB): 2.5–3 s with a graphics card, around 23 s without.

The booth fetches the large model by itself at start-up as soon as a template uses AI cut-out, and switches to it once it is there. If you pick a model yourself, that choice stands – on a weak machine the built-in one is the right answer.

If no time is shown: the model has never run on this booth. Open Tune AI once and let it process a test photo, then the number appears.

On iPad, iPhone and Mac, Apple Vision does the work right on the device. There is exactly one model there – nothing to choose, nothing to download, no graphics card needed.

Resolution: more is not better

Model resolution (on iPad, iPhone and Mac: Matte resolution) is set to 1024 and should stay there. A cut-out model is trained on a particular image size; above it, the model no longer sees the person as a whole but only fragments – and the mask falls apart. Too small is wrong as well: at 512 the built-in model lost an entire shoulder on a 30-megapixel photo. Models with a fixed input size hide the setting, because it would have no effect there.

Snap edge to image has been on by default since version 1.2.0: the booth traces the edge of the AI mask along the original photo – by color, and only where the mask is uncertain.

Order: cut out first, filter afterwards

Beauty and look are applied after the cut-out, onto the finished picture. That is not an accident: despill works with real green, and the edge refinement needs the colour image to find the hair edge. A black-and-white filter beforehand takes exactly that away. As a result the look sits on the subject and the background, which is more coherent anyway.

When nobody is in the picture

From version 1.7.0: if the AI cut-out finds nobody in the finished photo – for example because the guest stepped out of the frame when the shutter fired – the photo shows only the template's background. On the result screen the guest reads the reason:

We couldn't see anyone in the picture – the photo shows only the backdrop. Step in front of the camera and try again.

Result screen: the photo shows only the background, with the notice for the guest above it
Result screen: the photo shows only the background, with the notice for the guest above it

This applies to the AI cut-out of photos; with the green cloth the notice does not appear. If the cut-out fails because of a technical error, the guest gets their unaltered photo as before.

You change the wording of the notice under Design → Guest screens → Result, in the section Error texts, field No one detected in the picture (AI cutout) – also per guest language. If the field stays empty, the default text applies. It only appears if there is a template with AI cut-out.

When the result isn't right

  • Green fringe on hair: raise Spill strength (%), and if needed Edge choke (px) by 1–2 px.
  • Holes in the subject: tolerance too high, or the cloth is lit unevenly. Check the light before you touch sliders.
  • Edge looks cut with scissors: raise the edge softness.
  • AI loses body parts: check that the resolution is 1024; with dark clothing against a dark background, more light on the subject helps more than any setting.