AI Outfitting App

The brief

The client wanted an application that would allow anyone to try on outfits instantly. The user uploads their photos and of their clothes, and generates the look on themselves, with whichever combination they like. An AI assistant also assists with recommendations based on their wardrobe and preferences.

The product had to work from photographs taken on a phone, in a bedroom, in whatever light was available. Anything that required good photography would not survive contact with real users.

The challenge

Three problems had to be solved together, and each one constrained the others.

• Recognising garments from casual photographs - uneven light, cluttered backgrounds, items folded or on a hanger rather than worn
• Matching by colour and style in a way that reads as taste rather than arithmetic, since a technically correct pairing can still look wrong
• Staying inside a latency budget that keeps a mobile app feeling immediate, with inference that cannot always run on the device

How we approached it

We split the problem into a pipeline rather than trying to solve it with one model. Each stage does something narrow and testable, which made it possible to improve accuracy without retraining everything.

Garment segmentation isolates each item from its background. Attribute extraction reads colour, pattern, material cues, and garment type. The next stage then assembles combinations, with a generative layer handling occasion and context. An image model is then used to fit the item to the person's body.

What we built

A cross-platform client covering web, iOS, and Android from one codebase, backed by an inference service that keeps the heavier models off the device.

Each user's wardrobe is indexed as it is photographed, so recommendations are drawn from a structured view of what they own rather than re-analysed on every request.

Seamless experience where the user can quickly try on different outfits and combintations, as well as get advice on what to wear for a given occasion, all from their own digital (uploaded and saved) or physical wardrobe.

Results

The figures below are the client's own measurements taken after launch.

After launch

Recommendation quality is the kind of thing that improves with usage data, so the pipeline was built to be retrained on real generations and interactions rather than frozen at launch. The client owns the codebase.