The problem
Off-the-shelf AI tools were poor and expensive. A renter receives the exact garment in the photo, and image models like to improve what they see: in a live test, one redesigned a dress.
What I built
- A product workspace: generate, approve, push, with preview and undo.
- An accuracy system: the ghost mannequin renders first and a vision QC judge checks garment, pose and outfit.
- One licensed model per product, plus styled looks.
Key decisions
- Fidelity before polish: the output garment must match the references.
- Never invent a view: side or back shots only when a reference exists.
- No silent store writes: preview, confirm, log, undo.
The hardest problem
Only about 15% of products looked like different moments of a shoot. I split what must never change (garment, body angle, identity) from what should (head, expression, hands, stance), then varied the front shots, at no extra model cost.
Results
- Interface review score 18 → 31 out of 40.
- About 30 min by hand¹ → 47 s median render.
¹ My estimate.