CatVTON is a diffusion model for virtual try-on, designed for efficiency and ease of use. It targets researchers and developers in computer vision and fashion tech, offering a lightweight architecture for high-resolution image generation with reduced VRAM requirements.

How It Works

CatVTON leverages a diffusion model architecture, specifically building upon Stable Diffusion v1.5. Its novelty lies in a "concatenation" approach, enabling parameter-efficient training and simplified inference. This method allows for a total network size of 899.06M parameters, with only 49.57M trainable, and inference requiring less than 8GB VRAM for 1024x768 resolution.

Quick Start & Requirements

• Install: pip install -r requirements.txt within a conda environment.
• Prerequisites: Python 3.9.0, CUDA. Datasets like VITON-HD or DressCode are required for inference.
• Deployment: ComfyUI workflow and Gradio app are available.
• Docs: https://arxiv.org/abs/2407.15886

Highlighted Details

• Accepted to ICLR 2025.
• Supports 1024x768 resolution with < 8GB VRAM.
• Parameter-efficient training (49.57M trainable parameters).
• Mask-free version available.
• Integrates with ComfyUI.

Maintenance & Community

• Active development with recent updates (CatV2TON, FLUX.1-Fill-dev LoRA).
• HuggingFace Space available.
• Maintains an "Awesome-Try-On-Models" repository.

Licensing & Compatibility

• Licensed under Creative Commons BY-NC-SA 4.0.
• Non-commercial use only. Contributions must be shared under the same license.

Limitations & Caveats

The project is primarily tested on Linux; Windows users may encounter issues (refer to issue#8). The Gradio app is noted as not a stable version.