How I built it

Vlad Ryzhkov · November 2026

Clothes get photographed where they hang: on a rail, a hook, a mannequin, or on a person. Background removers handle the background. They don’t know the hanger isn’t part of the garment, that an arm is hiding half the shirt, or where the hair ends and the shop begins.

Cutting is solved. The part I built decides which piece is the garment. The rest is off the shelf:

  1. Find every piece. SAM 3 and BiRefNet.
  2. Decide what belongs to the garment. Claude Haiku judges each doubtful piece.
  3. Rebuild what was hidden. Ideogram’s edit model redraws what a hanger or an arm covered, then it cuts again.
  4. Clean the edges. ViTMatte and FBA matting, for knit, tulle, fur and hair.

One rule runs through it: a damaged garment is worse than a leftover hanger. So nothing is cut on one opinion. Every piece gets a second look first.

Hair got its own round. A flyaway strand is a few pixels wide and half see-through, with the street showing through it. The edge step keeps the strand and lets the street go, so the person comes out whole. It holds for flyaway hair and fur worn on a person; wigs on a stand are still hit and miss.

How I know it works

I keep 34 hard photos and 15 ordinary ones, and check every change side by side on the same photos. The current version against my first: 29 better, 5 the same, 0 worse. The 15 ordinary photos came out unchanged.

Running it

One L4 GPU that goes to sleep after about 5 minutes with nothing to do. The first photo after that takes a minute or two, then about 12 seconds a photo, roughly a cent each with everything in. This site sends at most 4 photos at a time, in turns, so one big batch never blocks anyone.

Nov 2026: hangers, mannequins, racks, hair.