Oxygen-TryOn: Fashion-Native Foundation Model for Any-item Virtual Try-On
Yong Liu, Xiaolong Fu, Zihang Xu, Wen Xue, Xueheng Li, Lin Song, Yuan Zhang, Chuyang Zhao, Haoyang Huang, Nan Duan, Yipeng Sun, Yan Li, Simiu Gu
cs.CV
2026-07-24
JD's Oxygen-TryOn reformulates virtual try-on as multi-reference understanding-driven generation, handles any fashion category and multi-item composition, and beats GPT-Image-2 and Seedream5 Lite across benchmarks.
Virtual try-on looks simple: given a product photo and a person photo, synthesize that person wearing the product. Existing systems get stuck in the same place. They handle a single garment category, assume the reference is a clean studio flat-lay, and break the moment the reference is an in-the-wild worn photo, or the moment you want to compose a top, pants, shoes, bag, and hat at once.
The strongest systems (GPT-Image-2, Seedream5 Lite, Nano Banana Pro) are closed and are really general image editors pressed into try-on duty. On the open side, models like FLUX.2 still trail. Oxygen-TryOn starts from a different bet: rather than repurpose a general editor, build a foundation model that is try-on-native from the start.
The core move is changing the paradigm. Where prior work treats try-on as mask-based inpainting (mask the region on the person, then fill), Oxygen-TryOn reformulates it as a multi-reference, understanding-driven generation task. The model reasons about what each reference is, where it goes, and what occludes what, then generates.
The base is JoyAI-Image-Edit: Qwen3-VL-8B for understanding, Wan-2.1 VAE for the latent space, and a 16B multimodal diffusion transformer (MMDiT) for generation. The subject and reference images are encoded by a shared VAE into tokens laid out as one sequence, with a target region of Gaussian noise that gets denoised and reference regions holding clean latents as conditioning. A counterintuitive choice is that references get no segment or index embeddings. Instead the temporal axis of multimodal RoPE assigns non-overlapping intervals to the target and each reference, and references are bound by natural-language position in the instruction ("Picture 1", "Picture 2"). The model aligns references by reading the instruction, not by dedicated tokens.
Training is three stages. Continued pre-training (CPT) balances general corpus and try-on data 1:1 to inject domain knowledge without drifting from the base. Supervised fine-tuning (SFT) uses a conditional flow-matching loss with dynamic sequence packing for unequal lengths. Reinforcement learning (RL) combines a hybrid reward: an in-house 8B try-on reward model (fine-tuned on about 100k human preference pairs, with separate heads for item consistency, identity consistency, and overall quality) and a Gemini 3.1 Pro judge scoring six rubric dimensions, aggregated by harmonic mean within aspects and geometric mean between them. Policy optimization uses the DiffusionNFT paradigm.
A data engine filters 50M+ raw images across garments, shoes, bags, hats, glasses, and jewelry, and builds training pairs through four complementary pipelines (real product plus real person, real product plus synthetic person, synthetic product plus real person, fully synthetic). That is what justifies calling the model fashion-native.
On public benchmarks (DressCode, VITON-HD), Oxygen-TryOn brings paired FID down to 1.93 on DressCode, well below FastFit's 3.66 and CatVTON's 5.03; on VITON-HD it drops from FastFit's 6.86 to 3.94.
On the VLM-judged TStars-VTON (1-10 scale), single-item overall reaches 9.36, ahead of Seedream5 Lite (8.77) and GPT-Image-2 (8.34); multi-item overall is 8.41, ahead of GPT-Image-2 (7.90) and Seedream5 Lite (8.19), trailing only on item fidelity to Seedream5.
| Benchmark | Metric | Oxygen-TryOn | Comparison |
| DressCode paired | FID lower | 1.93 | FastFit 3.66 |
| VITON-HD paired | FID lower | 3.94 | FastFit 6.86 |
| TStars single-item | Overall | 9.36 | GPT-Image-2 8.34 |
| TStars multi-item | Overall | 8.41 | Seedream5 Lite 8.19 |
On the in-house Oxygen-TryOn Bench (1,000 real samples, judged by GPT-5) and a 985-sample human eval, usability rate is the standout advantage: 86.8% in the cloth-to-model setting versus 80.4% for GPT-Image-2 and 67.5% for FLUX.2. In human eval the overall score (3.55) narrowly beats GPT-Image-2 (3.54), with higher subject consistency (3.75 vs 3.64), while GPT-Image-2 edges ahead on item consistency and aesthetics.
A training-recipe ablation shows the big jump happens at SFT (usability rises from CPT's 68% to 86%); RL mostly sharpens fine-grained consistency, on the order of one point.
For e-commerce and content generation, this means virtual try-on can finally handle whole-outfit composition and real-person reference photos, the two scenarios that used to need manual cleanup, at usability rates above 85%, which is what matters for shipping at scale. Reframing try-on as understanding-driven multi-reference generation also gives follow-on work a clean paradigm: multiple references can be bound by natural-language position in the instruction, without dedicated tokens.
Honesty requires saying that on aesthetics and texture reproduction it trades blows with the strongest closed systems. A lead, not a rout.
The authors acknowledge that coordinating five or more references under complex layering gets hard, that multi-item item fidelity is bounded by the base model's capacity, and that some proprietary systems still reproduce texture better in multi-item settings.
Two things feel undertested. The RL stage gains only about one point, and relative to the engineering cost (a custom reward model plus a Gemini judge plus DiffusionNFT) the paper does not discuss whether that ROI holds. Most results also come from the in-house bench and re-evaluated public subsets; readers should check how those subsets were selected.