// Fashion & retail · Virtual try-on mobile app
A virtual fitting room, powered by an image model on private GPU hardware.
A mobile app that shows a shopper wearing an outfit from a single photo, generated by a self-hosted image model.
// The challenge
What stood in the way
Shoppers want to see how an outfit looks on them before they buy, from nothing more than a photo.
General-purpose image models tend to change the person along with the clothes, so faces, poses and backgrounds drift.
// The solution
How we solved it
We built a React Native app in which a user picks a photo of themselves and a garment, then gets back an image of themselves wearing it, with a history of past results to save and share.
Behind it, a FastAPI service runs a 20-billion-parameter image-editing model with a fast 8-step sampler on self-hosted GPU hardware, with no third-party image API. Garment backgrounds are removed first, and the model is instructed to keep face, pose and background unchanged so that only the clothing changes.
Admission control keeps the GPU responsive under load: work is limited to what the hardware can run, a short queue absorbs bursts, and anything beyond it is turned away cleanly with a retry hint instead of timing out.
// Results
- parameter image model, self-hosted
- 20B
- sampling steps per try-on image
- 8
// Start a conversation
Have a similar challenge?
Tell us about your systems and goals, and we'll show you how we would approach it.