How JD’s AI Try‑On “Oxygen Tryon” Revolutionizes Online Fashion Shopping

Introduction

Traditional online shopping often leaves consumers unable to see how clothes look on them, leading to high return rates. In September, JD unveiled the AI try‑on system “Oxygen Tryon,” which uses AI to generate realistic virtual try‑on images from a single photo, improving satisfaction and reducing returns.

Business Scenario

In fashion e‑commerce, return rates can reach 30‑40%, causing extra logistics costs, product loss, and customer‑service pressure. Reducing returns and improving conversion are critical challenges.

Technical Challenges

Since 2015, the industry has moved from GAN‑based static image synthesis to real‑time body tracking and AR‑based 2.5D models. Remaining difficulties include accurate body motion, realistic clothing physics, complex scene adaptation, and harmonious lighting.

Technical Practice

Oxygen Tryon first extracts human keypoints and creates a mask for the target clothing area. The masked image is processed by a Redux model to extract clothing features, which are fed as prompt_embed into the Fluxfill model to generate the final try‑on image.

Technical Innovations

Precise body recognition: JD’s proprietary algorithm accurately analyses body dimensions to ensure natural drape and fit.

Realistic material rendering: High‑fidelity rendering of fabrics such as cotton, silk, and denim captures light reflection, refraction, folds, and texture.

Fast generation: High‑quality try‑on images are produced within 7 seconds, providing a smooth user experience.

Intelligent outfit recommendation: A multimodal model suggests coordinated outfits and supports virtual try‑on.

Future Outlook

JD plans to expand Oxygen Tryon to over 30 brands and 100,000+ SKUs for the 11.11 shopping festival, improve model consistency for complex textures and logos, enhance physical simulation of fabrics, and add personalized features such as skin‑tone matching, size assistance, and support for shoes, jewelry, and accessories.