Oxygen-TryOn is a fashion-native model for multi-item virtual try-on

JD-company · hf · 2026-07-28

JD’s team presents Oxygen-TryOn, a fashion-native foundation model for any-item virtual try-on.

Unlike general-purpose image editors repurposed with prompts, the model is built specifically for try-on.
It accepts one or more reference items and a target person image, then synthesizes a photorealistic image of that person wearing the items.
The system supports a wide range of fashion categories, including clothing, outerwear, accessories, footwear, and bags.
It handles full-body and half-body views, variable numbers of references, and free multi-item composition.
The training stack includes a large-scale data engine plus a three-stage recipe: continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL).
The RL stage uses a hybrid reward combining an in-house try-on reward model and a rubric-guided general-purpose model.
The authors claim state-of-the-art consistency and realism on public benchmarks and their internal benchmark, including strong results on multi-item try-on versus systems such as Nano Banana Pro, GPT-Image-2, Seedream5 Lite, and FLUX.2.

More from Multimodal