TryOnDiffusion

TryOnDiffusion: two-stage diffusion VTON with Parallel-UNet for realistic, detail-preserving virtual try-on.

Fashion Updated 13 hours ago

Visit Website

TryOnDiffusion's Top Features

Diffusion-based architecture for virtual try-on

Two-stage pipeline: Warping Module and Try-On Diffusion Module

Parallel-UNet design to preserve garment details during synthesis

Zero-shot try-on from arbitrary person and in-shop garment images

Robust handling of pose variations and warping misalignments

High-fidelity garment detail preservation and realism

Trained and evaluated on DeepFashion and VITON datasets

Open-source code and pretrained models available

Validated with results and comparisons on the project website

Research published and presented at CVPR 2023

Frequently asked questions about TryOnDiffusion

TryOnDiffusion is a diffusion-based Virtual Try-On framework that creates realistic try-on images by combining a person photo with an in-shop clothing image in a zero-shot setting.

It replaces single-stage generators with a two-stage diffusion pipeline that improves warping realism and preserves garment fidelity, reducing artifacts in challenging poses.

A two-stage Warping and Try-On Diffusion process, a Parallel-UNet design, robust handling of pose misalignments, and strong garment detail preservation.

DeepFashion and VITON are used to train and evaluate the system across diverse clothing types and human poses.

Yes. The website links to the GitHub repository with code and trained models.

A team from the Institute of Automation, Chinese Academy of Sciences, and the University of Chinese Academy of Sciences.

First, a Warping Module aligns the garment with the target person’s pose; second, a Try-On Diffusion Module synthesizes the final image while preserving garment details.

Yes. The website provides links to supplementary materials and pretrained models.

Yes. It is designed to handle large pose variations, occlusions, and diverse clothing categories.

Yes. The website showcases qualitative results and comparisons with prior state-of-the-art, highlighting improved realism and garment fidelity.