Oxygen-TryOn Claims SOTA Virtual Try-On Over GPT-Image-2, FLUX.2

A new paper on arXiv introduces Oxygen-TryOn, described as a unified foundation model for any-item virtual try-on. What makes it worth reading rather than skimming: the authors say they got there not by repurposing a general image editor, but by treating try-on as its own problem class from the ground up.

The reframing is the part that stands out. Instead of the mask-based inpainting approach most try-on systems inherit from image editing, the paper describes try-on as a 'multi-reference, understanding-driven generation task.' In practice that means the model handles full- and half-body views, a variable number of reference images, and free multi-item composition without treating each as a special case, while trying to preserve subject identity and item appearance. Training runs in three stages, continued pre-training, supervised fine-tuning, and reinforcement learning, with the RL step powered by a hybrid reward that combines an in-house try-on reward model with a proprietary, rubric-guided general-purpose model.

Why this matters if you are not training generation models yourself: virtual try-on is one of the small handful of image tasks where fashion and commerce teams have been paying for hosted generalist APIs because the domain-specific models were not close. The arXiv abstract claims Oxygen-TryOn leads on multi-item try-on and matches or surpasses Nano Banana Pro, GPT-Image-2, Seedream5 Lite, and FLUX.2. If that holds up outside the authors' own evaluation, it is an argument that vertical foundation models can beat the horizontal frontier on specific commerce tasks.

The honest caveat: the abstract does not name the benchmarks it claims to lead on, does not disclose model size or cost per generation, and the RL stage relies partly on the team's own reward model, which is exactly the kind of setup where reported scores can drift ahead of what a third party would measure. Take the specifics as reported, not settled. What the reporting doesn't give you is availability either, whether the weights, an API, or neither will actually ship.

The direction is the part worth watching. If purpose-built fashion models can pull ahead of general image editors on their home turf, other visual verticals with similar shape have a template for doing the same.