Image try-on today

This part of the docs summarises the state of image virtual try-on as of September 2026. It draws on published papers, model cards, licence files and company disclosures. Where a figure could not be verified from a primary source, the text says so.

2.1The task

An image try-on model takes a photograph of a person and a photograph of a garment, and produces a photograph of that person wearing that garment. The garment photo may be a product shot or a photo of someone else wearing it. Two evaluation settings are standard. In the paired setting, the model re-dresses a person in the garment they already wear, so the output can be compared with the real photograph. In the unpaired setting the garment is new, so only distribution-level realism can be scored.

Models also differ in what else they need. Older methods require a mask of the region to repaint, a pose map and a body-part segmentation. Newer mask-free methods take only the two photographs, which makes them easier to deploy and less brittle when segmentation fails.

Video and live try-on. Video try-on takes a video of a person and a garment image, and returns the video with the person wearing the garment. It must keep identity, motion and background, and keep the garment identical from frame to frame. Live try-on does the same on a camera stream as it arrives. Each frame must be ready in about 33 to 66 milliseconds, drawn only from the current camera frame and what the model has already produced, because future frames do not exist yet.

2.2How the architectures evolved

OOTDiffusion, often cited as the reference open model, belongs to the 2024 generation. It pairs a Stable Diffusion 1.5 denoiser with an outfitting network whose features are fused through self-attention, and it reports results at 512×384 [25]. Its fusion idea survives in newer designs. Its backbone, resolution and CC BY-NC-SA 4.0 licence make it unsuitable as the foundation for a commercial model today.

2.3Reported results on the standard benchmark

Table 6 lists results on VITON-HD as reported in each paper. The numbers are indicative only. Papers differ in resolution, FID implementation and whether baselines were re-run, so differences below about one FID point are within that noise.

Two conclusions follow. First, VITON-HD is saturating: the spread among modern methods is small relative to measurement noise, and the benchmark itself covers only upper-body garments. Second, general image editors are now serious competitors. On VTEdit-Bench, published at ECCV 2026, FLUX.2 models ranked best overall among general editors and Qwen-Image-Edit-2511 came close [41]. In Alibaba's human study, Tstars-Tryon beat Google's Nano Banana Pro in 41.1% of comparisons, tied in 41.6% and lost 17.3% [3].

2.4What distinguishes the strongest systems

• Scale and synthetic pairs. Google trained TryOnDiffusion on about 4 million pairs. FASHN trained VTON 1.5 on 18 million masked pairs and then 4 million synthetic triplets, and JD built Oxygen-TryOn from more than 50 million raw images [31, 40, 15]. A synthetic triplet keeps a real photograph as the target and uses an earlier model to create the input. The model therefore learns to produce real photographs without needing a mask.
• Editing foundation models plus LoRA. Layering VTON fine-tuned Qwen-Image-Edit with LoRA of rank 32 on the attention projections, on one NVIDIA H200 GPU: 20,000 steps on 29,151 pairs, then 5,000 steps on 5,768 pairs [16]. TAMF-VTON reports the lowest unpaired FID in Table 6 with a mixture-of-experts LoRA on the same base [36].
• Preference optimisation. Tstars-Tryon and Oxygen-TryOn follow pre-training and supervised fine-tuning with reinforcement learning. Oxygen's reward combines an 8-billion-parameter vision-language model trained on 100,000 preference pairs with a Gemini rubric judge [15]. TryOnReward released about 100,000 human ratings for this purpose [42].
• Distillation and caching for speed. FastFit caches garment features for a 3.5× speed-up, DirectTryOn generates in one step in 0.48 seconds, and Tstars-Tryon serves in about 3.9 seconds after step distillation [38, 43, 3].
• Small specialists remain competitive. FASHN VTON 1.5 generates directly in pixel space with 972 million parameters and no autoencoder. FASHN states it can be trained from scratch for USD 5,000 to 10,000 [30, 44]. Pixel-space generation avoids the eight-fold compression that blurs fine print.

2.5Industrial landscape

Industrial try-on is now deployed across North America, Europe and Asia. This review found no public disclosure from these organisations of try-on quality broken down by skin tone or body shape, and no deployed generative try-on built for draped garments: Meesho's announced saree try-on relied on warping and 3D rendering rather than a generative try-on model [47].

2.6Open foundation models and their licences

For a company, the licence of the base model decides what can be sold. A LoRA inherits its base model's terms, and some licences also restrict using a model's outputs to train other models.

2.7Datasets and their licences

None of these datasets clearly permits commercial training. Evaluation on the public datasets is possible only where their terms allow research use. A commercial model must be trained on data that FabricVTON owns or licenses.

2.8How try-on is evaluated, and where evaluation falls short

FID and KID measure whether outputs look like real photographs, not whether the right garment was transferred. SSIM and LPIPS need a ground-truth photograph and are dominated by the background. OpenVTON-Bench measured agreement with human rankings at Kendall τ of 0.611 for SSIM, against 0.833 for its own metric combining a vision-language judge and segmentation [53].

Several try-on benchmarks appeared in 2025 and 2026: VTBench, VTONQA, OpenVTON-Bench, VTON-QBench with 431,800 human annotations, VTEdit-Bench and TryOnReward [54, 55, 56]. They improve on classic metrics, but none covers draped garments, and none reports results by skin tone or body shape. A new general-purpose benchmark would not be novel on its own. One that spans garments from many regions, includes real shoppers' photos, is stratified by skin tone and body shape, and is built from consented, licensed photographs would be.