关键词: 虚拟试衣, 扩散模型, 一致性, ControlNet

Abstract: "With the rapid advancement of image generation technology, diffusion models have emerged as a dominant approach in in the field of image synthesis and demonstrated significant potential across various application scenarios, with virtual try-on being one of its crucial application directions. Compared to traditional generative models, diffusion models excel in producing higher-quality images. However, due to their inherent randomness, it is challenging to ensure consistency between the generated try-on clothing and the reference clothing. Currently, despite extensive efforts to enhance consistency by incorporating conditional control into diffusion models, but they still face challenges such as accurately preserving clothing details and optimizing the integration quality between clothing and the human body. To address these issues, this paper proposes a multi-stage virtual try-on framework aimed at improving clothing consistency and enhancing the integration effect between clothing and the human body. The proposed method consists of two primary stages. In the first stage, a clothing deformation network integrated with a multi-scale feature attention module is designed to reduce texture distortion during the clothing deformation process, improve the stability of clothing deformation, and enhance the ability to preserve detailed textures. In the second stage, a local refinement network based on ControlNet is introduced to perform detailed refinement on the boundary regions between clothing and the human body. By incorporating dense pose mapping and clothing reference images as additional control signals, this network can effectively reconstruct clothing details, achieving a more natural integration of clothing with the human body. Experimental results demonstrate that the proposed method offers significant advantages in maintaining clothing consistency while improving the realism of the try-on images. Ablation experiments and comparative analysis with current mainstream virtual try-on frameworks further validate the superiority of this approach in preserving clothing details and optimizing the clothing deformation effect. The results indicate that this method has achieved significant improvements in enhancing the quality of virtual try-on, but there are still certain limitations. To streamline the workflow and improve overall try-on quality, future research will explore end-to-end training strategies and conduct training and testing on larger, more diverse datasets to enhance the model's generalization capability. By further optimizing the method and addressing existing challenges, this study aims to provide more reliable technical support for intelligent clothing design and digital clothing applications. The proposed method effectively tackles issues related to clothing deformation and detail preservation, significantly elevating the visual experience of virtual try-on and providing users with more realistic and reliable try-on effects. These improvements open up new possibilities for applications in intelligent clothing design, online shopping platforms, and digital clothing displays."

Key words: "virtual try-on, diffusion model, consistency, ControlNet