Hugging Face has introduced Oxygen-TryOn, a unified open-source foundation model for any-item virtual try-on. Unlike general-purpose image editors repurposed for the task, Oxygen-TryOn is built specifically for fashion try-on, using a dedicated data engine and try-on-specific training pipeline.
The model accepts one or more reference items—either clean product shots or in-the-wild photos of someone wearing them—along with a single target subject image, and synthesizes a photorealistic image of the subject wearing the items. It supports diverse fashion categories including clothing, outerwear, accessories, footwear, and bags, and can handle full- and half-body views, variable numbers of references, and free multi-item composition.
Oxygen-TryOn reformulates try-on as a multi-reference, understanding-driven generation task rather than mask-based inpainting. It is built on the JoyAI-Image-Edit architecture, coupling a multimodal large language model (MLLM) for reference and instruction understanding with a multimodal diffusion transformer (MMDiT) for high-fidelity synthesis. The training follows a three-stage recipe: continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL) with a hybrid reward combining an in-house try-on reward model and a proprietary rubric-guided general-purpose model.
The model also retains general instruction-based editing ability, such as pose changes, within the same generation pass. According to the paper, Oxygen-TryOn achieves state-of-the-art consistency and realism on single-item try-on and leads on multi-item try-on, surpassing both proprietary systems (Nano Banana Pro, GPT-Image-2, Seedream5 Lite) and open-source models (FLUX.2). It also demonstrates cross-domain generalization to stylized 3D avatars, illustrated characters, statues, and posters.
Oxygen-TryOn is claimed to be the first open-source system to deliver any-item, multi-reference try-on at this level of fidelity. The model is expected to be open-sourced, as indicated by a comment on the paper page.