tryoffdiff
TryOffDiff is an academic research repository implementing virtual try-on technology via high-fidelity garment reconstruction using diffusion models. Developed by Riza Velioglu and colleagues, this project combines two primary research initiatives: TryOffDiff for single-garment application and MGT for extending capabilities to multi-garment scenarios. The software leverages PyTorch, Hugging Face Diffusers for diffusion components, and Accelerate for multi-GPU training. It utilizes Stable Diffusion v1.4 as the base model and SigLIP as the image encoder to generate realistic images of models wearing specific clothing items. Key features include training scripts for model development, inference pipelines for image generation, evaluation tools using metrics like IQA and DISTS, and dataset handling for VITON-HD and Dress Code. The repository is organized following the Cookiecutter Data Science structure, providing Jupyter notebooks, configuration management, and visualization tools. Notably, the code has been acce