Technical approach

3.1Objectives

3.2Design principles

• Measure before modelling. The benchmark and evaluation harness come first, so every model decision is scored across garment families, photo conditions and diverse bodies.
• Commercially clean by construction. Only Apache-2.0 or MIT bases and teachers, licensed or consented data, and a provenance record for every training sample.
• Coverage and fairness, not raw scale. A small team cannot match 50 million images. It can build the most diverse and best-measured data for the garments, bodies and photos that large players' published results leave out.
• Large teacher, small student. Train the best model on a 20-billion-parameter base, then distil it into a model cheap enough to serve on L4 and L40S GPUs.
• Open science where it helps the field. Publish methods, the benchmark and the evaluation toolkit. Keep the production model and licensed training data proprietary.
• Image first, live video as the goal. The image model comes first because it teaches the video model: it re-dresses real clips to make training pairs. Every image-track decision is judged by whether it also serves the video track.

3.3System overview

Figure 1 shows how the seven work packages connect. The image track turns licensed data into a servable image model. The video and live track reuses that model as a teacher and ends in live try-on. The measurement track scores every checkpoint in both.

3.4WP1: Benchmark and fairness audit

The benchmark is a test set built from new, licensed photography and licensed partner catalogues. Each item pairs a photograph of a consenting model with a garment presented in one of four ways: flat or unfolded, folded as sold, on a hanger or mannequin, or worn by another person, with a back view wherever it differs from the front. This lets the benchmark measure how much the garment presentation drives failure, separately from the model. Person photographs come in two conditions, studio and real-shopper (phone selfies, mirror shots and dim light), so the benchmark also measures how much the photo drives failure. Subjects are balanced across all ten groups of the Monk Skin Tone scale and across body shapes, including plus sizes [24].

At least ten systems will be audited without retraining. Open specialists include FASHN VTON 1.5, CatVTON, IDM-VTON, Leffa, FitDiT and OmniTry. Open general editors include Qwen-Image-Edit-2511 and FLUX.2 klein base 4B. Commercial APIs, such as FASHN's API and Google's virtual-try-on-001, will be included where their terms allow evaluation. Checkpoints with non-commercial licences will be run only where their terms allow research evaluation, and their results will appear only in publications.

Every subject signs a consent form covering AI training, evaluation and publication, in line with India's Digital Personal Data Protection Act and, for subjects in the United Kingdom or the European Union, the UK GDPR and the GDPR. Faces are anonymised in the public release where consent does not cover publication, and the release carries a datasheet.

The benchmark also has a video hold-out of about 300 clips, with people and garments kept out of all training data: 100 studio clips, 100 consented phone clips, and 100 live-camera stress clips recorded on a webcam, with low light, fast spins, hands and bags across the body, back turns and sequences of a minute or more.

3.5WP2: Data engine

Every sample enters a provenance ledger that records its source, licence, consent reference and any teacher model. Filters remove low-quality and duplicate pairs, and each batch is spot-checked by hand. The target is at least 20,000 training pairs, balanced across garment families and subjects. Generating 500,000 candidate synthetic triplets with the current teacher would cost about USD 2,000, based on the measured 6.5 seconds per try-on on an L40S billed at USD 2.27 per hour [7].

3.6WP3: Model adaptation

Base model

Qwen-Image-Edit-2511 is the base model for four reasons. It is licensed Apache-2.0, with no restriction on outputs. It accepts several input images natively, so the person and each garment view can be supplied separately. It is the base used by strong 2026 try-on work, including Layering VTON at ECCV 2026. And it is supported by the main open training tools, including musubi-tuner [48, 16, 61]. FLUX.2 klein base 4B, also Apache-2.0, is kept as the student model for serving.

Conditioning on several views of one garment

The central modelling idea addresses the ill-posed inputs described in the problem chapter. Instead of one garment photograph, the model receives several views of the same garment: the front and the back, a close-up of any print, logo or embroidery, and, for a draped or wrapped garment, the unfolded fabric and its decorated border or end. For a multi-piece outfit it receives each piece. A short instruction names the garment type and how it is worn, for example a coat worn open over a shirt, or a saree in a Nivi drape. The research question is whether this multi-view conditioning, together with the wear instruction, recovers detail that a single product photograph cannot supply.

Training recipe

3.7WP4: Preference tuning and distillation

Once supervised fine-tuning works, a panel of raters from several regions will compare outputs in pairs. Their judgements will train a small reward model or calibrate a vision-language judge, which then guides reinforcement learning of the kind used by Tstars-Tryon and Oxygen-TryOn [3, 15]. This stage improves the product and is optional for the papers.

A 20-billion-parameter model is expensive to serve. The programme will first test the Apache-2.0 four-step Lightning LoRA for Qwen-Image-Edit-2511, then distil the fine-tuned model into FLUX.2 klein base 4B [62]. The target is a distilled model within twice today's measured 6.5 seconds on an L40S, measured with FabricVTON's existing speed laboratory. The same work package replaces a third-party component in the current serving pipeline with a commercially licensed one.