Texture-Preserving Diffusion Models for High-Fidelity Virtual Try-On

Our Texture-Preserving Diffusion (TPD) model generates high-fidelity virtual try-on images without using specialized garment image encoders. It leverages the self-attention blocks in the diffusion model's denoising UNet to efficiently transfer textures from the reference garment to the person image. Additionally, it predicts an accurate inpainting mask to preserve the background and body details.