基于外观流和扩散模型的虚拟试衣  

Virtual Try-on Based on Appearance Flow and Diffusion Model

作　　者：陈相宜 CHEN Xiangyi(The School of Information Science and Technology,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]西南交通大学信息科学与技术学院,四川成都610031

出　　处：《现代信息科技》2025年第5期17-24,共8页Modern Information Technology

摘　　要：文章调研了国内外在虚拟试衣领域的研究现状,包括基于对抗生成网络和扩散模型等技术的应用情况,并总结了虚拟试衣各数据集。在相关理论基础上,对基于外观流和扩散模型的虚拟试衣进行研究。首先运用基于外观流的局部流全局解析(Local-Flow Global-Parsing,LFGP)模型对目标服装进行变形,使其符合人体姿势;接着使用扩散模型对人物和服装进行合成,并对细节进行处理,完成最终试衣。定性和定量分析实验结果表明,相较于以往研究,基于LFGP的服装变形可以更好地处理服装细节以及人物与服装的连接部分,扩散模型在试衣合成处理时也可以在保持真实性的同时提高灵活度,为虚拟试衣技术的进一步研究提供参考。This paper investigates the research status of virtual try-on field at home and abroad,including the application situation of technologies based on Generative Adversarial Network and Diffusion Model,and summarizes various datasets of virtual try-on.Based on related theories,it studies the virtual try-on based on appearance flow and Diffusion Model.Firstly,the Local-Flow Global-Parsing(LFGP)model is used to transform the target clothing to conform to the human posture.Then,the Diffusion Model is used to synthesize the characters and clothing,and the details are processed to complete the final tryon.Qualitative and quantitative analysis of the experimental results show that compared with previous studies,clothing transformation based on LFGP can better handle the clothing details and the connection parts between characters and clothing,and the Diffusion Model can also improve flexibility during the try-on synthesis processing while maintaining the authenticity,providing reference for further research of virtual try-on technologies.

关 键 词：虚拟试衣 深度学习 扩散模型 外观流 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]