Three screens for the shopper. Fourteen subsystems behind them. Here is what actually happens in those nine seconds.
Format, size, and whether the photo is actually usable — one person, front-facing, sharp enough. Rejected in under half a second, at no cost.
Input gates run before any AI spend: NSFW, apparent minors, policy. Output is checked again before the shopper sees anything.
Normalise orientation and colour, person-aware crop, and resolve the garment to a phrase like "a red plaid flannel shirt" — cached per product.
Pick a model that has cleared the quality benchmark, then optimise for cost and latency within that pool. Never the other way round.
The person photo, the garment, and a versioned prompt that names three constraints: preserve identity, realistic fit, keep background.
Identity drift, garment presence, artefacts, NSFW. A failure fails over to another provider rather than shipping a bad image.
Watermark, EU AI-disclosure badge, and machine-readable provenance metadata. Stored on a TTL-bucketed path behind signed URLs.
The playground runs this exact pipeline — the same one API consumers hit.
Open the playground