Executive summary

FabricVTON is an early-stage Indian company building photorealistic virtual try-on for fashion e-commerce worldwide. This programme has one end goal: live video try-on, in which shoppers see themselves wearing a garment on their own camera, in real time, for any garment and any body.

0.1The programme

The research runs for eighteen months. It builds image try-on first, because the image model becomes the teacher for the video model. Along the way it closes a gap the research community has left open: try-on that works equally well for every garment and every shopper, whatever their skin tone, body shape or region.

0.2The problem

About one in four online apparel orders in the United States is returned, and size and fit is the most-cited reason [1]. Virtual try-on is the industry's response, and it is now mainstream across markets: Google offers it in Search in the United States, the United Kingdom and India, Inditex runs Zara Try-On in 43 markets, and Alibaba serves it to millions of Taobao users [2, 3, 12, 13].

Yet research models are trained and scored almost entirely on a narrow wardrobe of stitched studio garments, from datasets whose terms forbid commercial use [4, 5], and no published work audits them by skin tone or body shape. Live try-on has only just become possible, and only on closed engines: the leading live try-on model costs USD 1.20 a minute, about a hundred times the cost of a photo try-on on FabricVTON's own infrastructure [7, 70]. The one open research system for live video try-on runs at 512×384 and has not released its code [71].

0.3Where FabricVTON stands

FabricVTON already operates a production photo try-on service on its own GPU infrastructure, built on an open, permissively licensed try-on model. On that infrastructure a photo try-on takes about 6.5 seconds on an NVIDIA L40S GPU and costs about USD 0.019 on an NVIDIA L4 [7]. It has also built a live try-on page with a consent screen, a session time limit and one-time session tokens, currently running on a licensed third-party engine. The programme therefore starts with a working baseline, a measurement harness, a permissively licensed teacher model and a live shell ready for its own engine.

0.4The approach

The programme has two tracks.

• The image trackbuilds a licensed and consented benchmark of at least 2,000 try-on pairs across at least eight garment families from at least five world regions, audits at least ten current systems for fairness, and fine-tunes FabricVTON's own image model on Qwen-Image-Edit-2511, an Apache-2.0 image editor.
• The video track, the end goal, uses that image model as a teacher: it re-dresses real, consented video clips to make training pairs, trains a video try-on model on the Apache-2.0 Wan2.1 video model, and then converts it into a live model that draws each frame from the camera frame, a memory of the garment and its own recent frames, at 15 or more frames per second on one GPU.

0.5Outputs

Four peer-reviewed papers, a public benchmark with image and video splits, an open evaluation toolkit, commercially usable image and video try-on models, and live try-on on FabricVTON's own model at an estimated tenth of the rented engine's cost or less. Compute is estimated at USD 60,000 to 105,000. Funding would also pay for photo and video data, human evaluation and research personnel, none of which the company can fund today. See What funding enables.

0.6At a glance