Oxygen-TryOn: A Fashion-Native Foundation Model for Any-Item Virtual Try-On

Authors: 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

Submitted: 23 July 2026

Abstract

We present Oxygen-TryOn, a unified foundation model for any-item virtual try-on. Unlike prior approaches that adapt general-purpose image editors for try-on tasks, Oxygen-TryOn is fashion-native—it is built specifically for virtual try-on through a dedicated data engine and try-on-specific training. Given one or more reference items (either clean product shots or in-the-wild worn-on photos) and a single target subject image, the model synthesizes a photorealistic image of the subject wearing those items across virtually any fashion category.

Existing systems are typically limited to a single garment category in a studio setting, and recent multi-reference methods remain garment-centric. In contrast, Oxygen-TryOn supports diverse items and scenarios, including full- and half-body views, a variable number of references, and free multi-item composition, while faithfully preserving both subject identity and item appearance.

Technical Approach

Rather than relying on mask-based inpainting, we reformulate try-on as a multi-reference, understanding-driven generation task. Our approach is built around three key components:

1. Data Engine

We developed a large-scale data engine that collects, manufactures, annotates, and filters high-quality try-on data at an unprecedented scale.

2. Three-Stage Training Recipe

Oxygen-TryOn uses a three-stage training pipeline:

• Continued Pre-Training (CPT): Foundational learning on large-scale fashion data.
• Supervised Fine-Tuning (SFT): Task-specific adaptation using curated try-on pairs.
• Reinforcement Learning (RL): A hybrid reward mechanism combines an in-house try-on reward model with a proprietary, rubric-guided general-purpose model. This jointly supervises fine-grained consistency and instruction-level quality, enabling the model to follow general editing instructions (e.g., pose changes) in the same pass.

3. Versatility and Performance

Across public benchmarks and our in-house Oxygen-TryOn Bench, the model achieves state-of-the-art consistency and realism on single-item try-on and leads on multi-item try-on. It matches or surpasses both leading proprietary systems (e.g., Nano Banana Pro, GPT-Image-2, Seedream5 Lite) and open-source models (e.g., FLUX.2).

Broader Context (2026)

As of 2026, the virtual try-on landscape has evolved rapidly, with e-commerce and augmented reality platforms demanding high-fidelity, real-time generation across all garment types and body poses. Oxygen-TryOn addresses this need by unifying previously fragmented approaches into a single, scalable foundation model. Its ability to handle multiple reference images, varied clothing categories, and natural pose variations makes it particularly suitable for real-world applications in online retail, virtual fitting rooms, and personalized fashion recommendations.

via ArXiv CV