CatVTON — virtual clothing try-on

★ 1.9K

CatVTON is an open-source virtual try-on diffusion model (accepted at ICLR 2025): simple and lightweight, with 899M parameters in total and 49.57M trainable. Reach for it when you need to show how a garment looks on a specific person without a photo shoot, for e-commerce product cards, lookbooks and fitting rooms; when try-on must run locally without uploading photos to a cloud; or when GPU memory is tight. How to use it: a conda environment with Python 3.9 and pip install -r requirements.txt, the Gradio app app.py, a ComfyUI workflow, or inference.py for the VITON-HD and DressCode datasets; weights download from Hugging Face automatically. In bf16, generating at 1024x768 needs about 8 GB of VRAM. CatVTON-FLUX is an extremely light 37.4M LoRA for FLUX.1-Fill-dev, and a mask-free version also exists. For video try-on the authors point to a separate model, CatV2TON. The direction is putting clothing onto a photo of a person, not generating clothing from scratch. A CUDA-capable GPU is needed. The repository licence could not be detected automatically, so check its terms before any commercial use. Use only with the consent of the people whose photos are processed.