Multi-View Clothing Images Generation Based on VAE and Diffusion Model

The generation of multi-view clothing images is more challenging due to the high deformability and rich details of clothing. In this paper, a multi-view clothing images generation network based on variational autoencoder (VAE) and diffusion model is proposed, which adopt a coarse-to-fine strategy to generate clothing images with different viewpoints from a single input image. First, fuzzy images with overall appearance structures are generated with VAE, which obtain the clothing's global structure and contours, ensuring consistency across different views. Then, the blurred images are fed into a diffusion model with deep U-Net structure for refinement, which significantly improves the clarity and detail recovery of the image. The proposed method has been extensively experimented on a large multi-view clothing image dataset MVC, and the results show that, compared with existing methods, our method generates more detailed multi-view clothing images with better consistent overall appearance and better visual perception.