Virtual Try-On: AI-powered virtual dressing

When buying clothes online, customers decide without seeing how the garment will look on them. A retailer's product photos are not uniform either: some on a mannequin, some on a hanger, some laid flat, some on a model.

The goal was to turn these mixed product photos into a dressing-ready catalogue, show the product realistically on the customer's own photo, and offer it as a service that plugs into the retailer's existing website.

Product extraction prepares two images per garment: an exact pixel cut-out from the photo and a clean product image of the garment on its own (parts hidden behind arms, hair or the mannequin are completed). Dressing uses the pixel cut-out, which gave the most faithful result in an internal blind comparison; if the cut-out fails inspection, the clean image is used.

For scene changes the person and the garment are separated into layers, the background is regenerated, a shadow is added under the feet and the person is scaled to the scene's perspective. Light matching is optional: when on, the scene lighting is applied to the person; when off, the person's image is left untouched.

Every result goes through automatic inspection. The user first sees an uninspected preview, and the final result arrives after inspection. A correction request changes only the relevant garment area in the selected version.

To connect a store's site there are endpoints for querying the library, matching by photo, starting a job and fetching the result; preparation can start as soon as the customer's photo is selected to shorten the wait.

The platform is live, running in Mipo's internal use and with client accounts. Duration, error rate and cost of every job are tracked in the admin panel; development continues.

In the fast generation setting, image generation dropped from 18 to 11 seconds. In a blind comparison of eight jobs by three judges, this setting produced the result most often ranked first (internal experiment, October 2026).