6.1Publications
Every Garment, Every Body: A Global Benchmark and Fairness Audit of Virtual Try-On
CVPR 2027 workshop; NeurIPS 2027 Evaluations and Datasets
Month 6 to 7Any Garment, Any View: Multi-View Pseudo-Paired Adaptation of Image-Editing Models for Virtual Try-On
ACM Multimedia 2027, BMVC 2027 or WACV 2028
Month 10Taught by Photos: Commercially Clean Synthetic Pairs for Garment-Faithful Video Try-On
CVPR 2028
Month 14Live Try-On: Garment-Faithful Streaming Virtual Try-On on One GPU
ECCV 2028 or SIGGRAPH 2028
Month 17 to 18| Output | Working title | Contribution | Target venue | Month |
|---|---|---|---|---|
| Paper 1 | Every Garment, Every Body: A Global Benchmark and Fairness Audit of Virtual Try-On | First licensed try-on benchmark spanning garment families from several regions, with garment attributes, photo conditions and fairness strata; audit of at least ten systems; judge calibration study | CVPR 2027 workshop; NeurIPS 2027 Evaluations and Datasets | 6 to 7 |
| Paper 2 | Any Garment, Any View: Multi-View Pseudo-Paired Adaptation of Image-Editing Models for Virtual Try-On | Data engine for garments beyond stitched studio wear; multi-view garment conditioning; distilled student model | ACM Multimedia 2027, BMVC 2027 or WACV 2028 | 10 |
| Paper 3 | Taught by Photos: Commercially Clean Synthetic Pairs for Garment-Faithful Video Try-On | Video data engine in which the image model re-dresses real clips; multi-view garment memory; a print and logo stability measure for video | CVPR 2028 | 14 |
| Paper 4 | Live Try-On: Garment-Faithful Streaming Virtual Try-On on One GPU | Causal student with persistent garment memory and self-forcing training; a live-camera benchmark; cost per stream-minute | ECCV 2028 or SIGGRAPH 2028 | 17 to 18 |
| Technical report (optional) | Cost-Aware Serving of Virtual Try-On on Commodity GPUs | Quality, latency and cost trade-offs for photo and live try-on, with blind human evaluation | arXiv; WACV applications track | 4 to 6 |
6.2Datasets and benchmark
- Global try-on benchmark. At least 2,000 consented test pairs across at least eight garment families from at least five world regions, with garment attribute labels, garment-presentation and photo-condition variants, and skin-tone and body-shape strata. Released for research with a datasheet.
- Human rating set. Pairwise preferences and attribute judgements from raters in several regions, released to support calibration of automatic judges.
- Proprietary training corpus. At least 20,000 licensed and synthetic pairs with a full provenance ledger. Kept private because licences and consents cover training, not redistribution.
- Video hold-out. About 300 consented clips in studio, phone and live-camera conditions, with people and garments kept out of training. Released for research with a datasheet.
- Proprietary video corpus. At least 7,000 real clips and 20,000 verified synthetic video pairs with a full provenance ledger. Kept private for the same reason as the image corpus.
6.3Models
- Research model. A LoRA fine-tune of Qwen-Image-Edit-2511 for try-on of any garment, used for Paper 2.
- Production model. A distilled student on FLUX.2 klein base 4B, deployed in FabricVTON's service once it beats the current model on the benchmark.
- Video try-on model. An offline model on Wan2.1 14B for garment-faithful video try-on, used for Paper 3 and as the live teacher.
- Live try-on model. A 1.3B causal student served over WebRTC behind FabricVTON's existing live page, replacing the rented engine once it wins on the hold-out.
- Baseline results. Published scores for at least ten existing systems on the benchmark.
6.4Software
- Evaluation toolkit, released open source under Apache-2.0, with the garment-attribute, skin-tone drift and body-distortion measures.
- Judge calibration scripts that measure agreement between automatic judges and human raters.
- Speed and cost laboratory for measuring latency and cost per try-on on cloud GPUs.
- Live evaluation harness that measures time to first frame, sustained frame rate, frame latency and drift on the live-camera stress clips.
6.5Key performance indicators
| Indicator | Target | Evidence | Month |
|---|---|---|---|
| Benchmark size | At least 2,000 image test pairs across at least eight garment families from at least five world regions | Datasheet | 5 |
| Fairness coverage | Subjects balanced across all ten Monk Skin Tone groups and across body shapes, including plus sizes | Datasheet | 5 |
| Systems audited | At least ten | Paper 1 | 6 |
| Rater agreement | Krippendorff's alpha of at least 0.6 on garment attributes | Pilot report | 2 |
| Image model quality | Statistically significant human-preference win over the best open baseline on garment fidelity | Paper 2 | 9 |
| No regression | No significant loss on standard metrics against the current production model | Evaluation report | 9 |
| Image latency | Distilled model at or below 13 seconds on an L40S | Speed laboratory report | 11 |
| Video data | At least 7,000 real clips and 20,000 verified synthetic pairs with provenance | Data card | 11 |
| Video model quality | Beats open video baselines on garment fidelity over time on the 300-clip hold-out | Paper 3 | 13 |
| Live speed | At least 15 fps at 512p on one GPU, with the first frame in under 2 seconds | Live evaluation report | 16 |
| Live cost | Under USD 0.12 per stream-minute, a tenth of the rented engine | Serving report | 16 |
| Live stability | Prints and logos stable across 60-second sessions by text accuracy and garment similarity | Live evaluation report | 16 |
| Publications | Two image-track papers submitted by month 10, and two video-track papers by month 18 | Submission records | 18 |
| Open release | Benchmark, video hold-out and evaluation toolkit public | Repositories | 12 and 18 |
6.6Intellectual property and open science
Methods, benchmark results and the evaluation toolkit will be published openly, with every contributor credited as a co-author. The benchmark test set and the video hold-out will be released under a research licence with datasheets. The production model, the training corpus and the provenance ledger remain FabricVTON's property, because the underlying licences and consents cover training but not redistribution. Any release of research weights will be decided per licence and consent, and stated in Paper 2.