TryOnDiffusion

TryOnDiffusion is a research framework for virtual try-on (VTON) that generates realistic images of a person wearing a specified garment, even when the person and garment come from different source images. Developed by researchers at the University of Washington and Google Research, it was presented at CVPR 2023. The system uses a two-stage diffusion process: a warping stage followed by a try-on diffusion stage, both built around a novel Parallel-UNet architecture. This design allows the model to implicitly warp the garment to match the target person's pose and body shape while preserving fine clothing details, all within a unified network rather than separate sequential steps. The framework is designed for zero-shot operation, meaning it can handle arbitrary person and garment images without requiring per-subject training. It processes inputs at multiple resolutions (128×128, 256×256, and up to 1024×1024 via super-resolution) to produce high-fidelity outputs. The project provides open-source code and pretrained models, and its results have been validated on standard benchmarks like DeepFashion and VITON. As a research project, it is not a commercial SaaS product; the website indicates a contact-based pricing model, which likely refers to research collaboration or licensing inquiries rather than a standard subscription.