Nexa VTO
Photorealistic virtual try-on from a single uploaded photo
The problem
Apparel returns are driven overwhelmingly by uncertainty about fit and look. A shopper looking at a garment on a studio model cannot tell how it will sit on their own body, so they order two sizes, keep one, and the retailer pays shipping both ways plus the cost of processing the return.
In store, the same uncertainty shows up as changing-room queues and abandoned baskets.
What we built
A virtual fitting platform. A shopper uploads a photo or captures one from their webcam and gets back a photorealistic image of themselves wearing the garment, with face, body shape, skin tone and pose preserved. It handles tops, bottoms and one-pieces such as dresses and jumpsuits.
Around the model sit the things that make it usable rather than a demo: automatic background removal, intelligent person auto-crop, and pose validation that rejects a bad input before inference rather than making the shopper wait for a result that was never going to work. Results are upscaled for HD output, kept in a per-session history with downloadable images, and browsable through a garment gallery filtered by gender and category.
It ships as a storefront and as a REST API, so a retailer can use the front end or embed try-on directly into their own product pages. A garment extraction utility onboards new SKUs straight from existing product photography.
How it is built
Vision models
- FASHN VTO v1.5
- IDM-VTON diffusion
- DWPose pose estimation
- DensePose body mapping
- RealESRGAN super-resolution
- rembg segmentation
Application
- FastAPI
- Next.js storefront
- REST API for embedding
Runtime
- Automatic CPU and GPU selection
- CUDA acceleration
Built by the same engineering team as the rest of the Nexa Suite.
Engineering decisions that mattered
Bad inputs rejected before inference, not after
Pose validation checks the uploaded photo is usable before any generation runs. The alternative is making a shopper wait for a result that was never going to work, which is where most try-on demos lose people.
Onboarding from existing product photography
A garment extraction utility takes new SKUs straight from the product photos a retailer already has, so adding a catalogue does not require a separate 3D or studio process.
Storefront and API from the same service
The same backend serves a ready-made storefront and a REST API for embedding into existing product pages, so a retailer can pilot without touching their site and integrate later without a rebuild.
How it is sold
Our own product, sold two ways: an on-premise deployment for retailers who need catalogue imagery to stay inside their own infrastructure, and a managed service for everyone else.
Target buyers run from e-commerce and D2C fashion brands through marketplaces, in-store kiosks, bridal and occasion wear, uniform and workwear suppliers, and rental and subscription services.
Common questions
- What does the shopper have to provide?
- One photo, either uploaded or captured from a webcam. Background removal and person cropping are automatic, and pose validation rejects an unusable photo up front rather than returning a poor result.
- Can it be embedded in our existing product pages?
- Yes. Alongside the ready-made storefront there is a REST API covering try-on, garments and results, so try-on can sit inside your own product pages. Retailers commonly pilot on the storefront first and embed once the catalogue and conversion numbers look right.
- Does our catalogue imagery have to leave our infrastructure?
- Not if that is a requirement. There is an on-premise deployment for retailers whose product imagery cannot leave their own network, alongside the managed service.
- Which garment types does it support?
- Tops, bottoms, and one-pieces such as dresses and jumpsuits. Fit and drape are hardest on loose and heavily structured garments, which is worth testing against your own catalogue rather than a sample set.
