Abstract: Image-based virtual try-on, as an economical and convenient technology in the virtual try-on domain, aims to synthesize realistic fitting effects by combining model images with clothing images. It has significant applications in online shopping, fashion design, and animation. Recently, generative large models represented by diffusion models have driven new breakthroughs and transformations in the field due to their stronger generative capabilities compared to traditional deep learning methods. However, there is a lack of comprehensive analysis and overview of image-based virtual try-on in the era of large models. This paper summarizes the key techniques of image-based virtual try-on, categorizes mainstream methods into three baseline processes: data preprocessing, warping generation, and try-on result generation. It also analyzes the implementation methods used in representative literature, compares major process methods, and introduces commonly used datasets, evaluation metrics, and loss functions in image-based virtual try-on. Finally, the paper discusses the challenges and limitations of image-based virtual try-on in the context of large models, and outlines future development and improvement directions for relevant technologies.

Key words: computer vision, virtual try-on, warping treatment, image synthesis, diffusion model

摘要： 基于图像的虚拟试衣作为虚拟试衣领域经济便利的一种技术形式，旨在通过模特图像与服装图像来合成逼真的试穿效果，其在网购、服装设计、动画等领域受到重点关注。近年来，以扩散模型为代表的生成式大模型凭借相比传统深度学习方法更强大的生成能力，推动了该领域的突破与变革。然而领域内缺乏对大模型时代下基于图像的虚拟试衣研究的进一步分析与全面概述。对基于图像的虚拟试衣进行汇总，按照数据预处理、翘曲生成和试穿结果生成这三步基线技术流程，对主流技术方法进行了划分和解析，对该领域代表性文献所用的实现方案进行了详细分析，并对主要流程方法进行了总结与对比。介绍了基于图像的虚拟试衣的常用数据集、评价指标与损失函数。最后结合所引的领域代表性文献，对大模型时代下基于图像的虚拟试衣存在的困难与不足进行了详细分析与分类，并据此对相关技术的未来发展与改进方向进行了概括与展望。

关键词: 计算机视觉, 虚拟试衣, 翘曲处理, 图像合成, 扩散模型