Overview
Most try-on demos are a website. TryOn is built as an engine: a multi-tenant FastAPI service with a versioned /v1 API and per-brand API keys, plus thin SDKs so a brand can put try-on inside its own Android app, iOS app or web store.
Generative try-on models render one fit regardless of size. TryOn renders the fit: size S sits tight, XL drapes loose, driven by the brand’s real size chart.
How it works
- Frames. The upload is sampled at about 4 fps, capped at 80 frames and downscaled to 720 px.
- Silhouettes. Each frame is thresholded and cleaned with morphology, keeping the largest connected component as the body.
- Camera model. The shopper turns in place, so the pipeline treats the capture as a turntable with evenly spaced viewing angles.
- Visual hull to mesh. A 128³ voxel grid is carved against every silhouette, then marching cubes turns the surviving volume into a
.glbbody mesh. - Size-accurate fit. A body or finished-measurement size chart, the garment’s cut and the shopper’s measurements give their true size and the drape and tension for any size they pick.
- 2D photo try-on.
POST /v1/tryon-photowarps a garment onto the person in their own photo, following their silhouette.
Testing, and its limits
- CI runs the API test suite, the 1,000-case placement stress test, and the web SDK, demo storefront and iOS SDK builds on every push.
- Body reconstruction is gated at 85% or better width-profile agreement with the source silhouette. That gate runs on synthetic bodies, not against professional 3D capture.
- Captures currently assume tight dark clothing on a lighter background and a slow in-place turn. Learned camera pose and physical cloth draping are next.
Stack
This project is not open source.