Virtual try-on that fuses outfits in latent diffusion space

What it does

OOTDiffusion is a research implementation for controllable virtual try-on: you feed it a photo of a person and a photo of a garment, and it renders the person wearing that garment. It handles both half-body shots (trained on VITON-HD) and full-body looks (trained on Dress Code), with separate categories for upper body, lower body, and dresses. The project ships with trained checkpoints, a Gradio demo hosted on Hugging Face, and inference scripts.

The interesting bit

Rather than warping or overlaying clothing onto a body in pixel space, the model fuses outfit features directly into the latent diffusion process. It also uses human parsing and pose estimation to keep limbs and fabric boundaries anatomically plausible.

Key highlights

• Supports both half-body and full-body garment transfer with category-aware inference
• Pre-trained checkpoints available for VITON-HD and Dress Code datasets
• Includes ONNX support for human parsing to reduce environment friction
• Live demo runs on Hugging Face Spaces with A100 GPUs
• Linux-only; authors have only validated on Ubuntu 22.04

Caveats

• Training code is still on the TODO list, so you can only run inference or study the architecture
• Requires manually downloading CLIP and several checkpoint files before it works
• Full-body mode needs explicit garment category flags (upper, lower, dress)

Verdict

Fashion-tech researchers and generative-image hackers who need a working virtual-try-on baseline will find the pre-trained weights useful. If you need end-to-end training code or Windows support, this is not your wardrobe yet.