丝绸2026，Vol.63Issue(1)：57-67,11.DOI:10.3969/j.issn.1001-7003.2026.01.007
深度学习在基于图像的二维虚拟试衣中的研究进展
Research progress of deep learning in image-based 2D virtual try-on
摘要
Abstract
With the rapid growth of online retail and demand for realistic try-on experiences,image-based 2D virtual try-on has become a research hotspot due to its low cost and strong adaptability.Recent advances in deep learning have significantly improved pose adaptation,detail fidelity,and image synthesis.From the interdisciplinary perspective of computer vision and fashion,this paper reviews the development of image-based 2D virtual try-on,summarizes the evolution of key technologies and performance comparisons,and discusses future directions,providing reference for further research and practical applications. Focusing on representative studies from 2017 to 2025,the paper first introduces the fundamental principles of generative adversarial networks(GANs),conditional GANs,and diffusion models,as well as the network architectures of basic virtual try-on models.Building on this,the paper conducts a systematic review around three core modules of virtual try-on—human pose estimation and feature extraction,clothing deformation,and image synthesis—supplemented by an integrated application module.Human pose estimation covers human segmentation,keypoint detection,dense pose,and contour representation,highlighting the role of feature fusion in improving pose adaptability and texture fidelity.Clothing deformation evolves from explicit geometric warping and flow-based deformation to implicit deformation integrating diffusion models and transformers,with implicit deformation emerging as the mainstream approach to enhance robustness.Image synthesis progresses from mask composition,convolutional neural networks,and 3D human-driven approaches to parser-free stages,culminating in a new synthesis phase that fuses diffusion models with transformers.To evaluate the performance of different techniques,representative models are quantitatively compared on the VITON-HD dataset,showing that diffusion-based networks generally outperform traditional GANs in image synthesis.Alongside improved generation quality,virtual try-on research has gradually expanded into integrated application scenarios,exploring user interaction,multi-layer clothing combination,and personalized recommendation cases. The contributions of this review are threefold.First,it extends the conventional three-stage workflow of pose estimation,garment deformation,and image synthesis by incorporating an integrated application stage,covering personalized recommendation and commercial deployment,thereby bridging research and practice.Second,it clarifies the technical transition from explicit to implicit deformation,and from mask-based composition to synthesis driven by diffusion models and transformers.Third,it summarizes the quantitative analyses of representative models,along with comparative evaluations of their strengths,limitations,and application scenarios,and highlights the advantages of diffusion-based methods in detail fidelity and pose generalization. Despite significant improvements in generation quality,existing models still face challenges in multi-layer clothing combinations,complex human poses,non-rigid deformation modeling,and real-time interaction,indicating limited generalization capabilities.Future research may advance high-quality virtual try-on through refined datasets,optimized model architectures,improved color accuracy,and closer integration of technology and industry.关键词
二维虚拟试衣/人体姿势/服装形变/图像合成/个性化推荐Key words
2D virtual try-on/human pose/garment deformation/image synthesis/personalized recommendation分类
轻工纺织引用本文复制引用
赵翎彤,徐军..深度学习在基于图像的二维虚拟试衣中的研究进展[J].丝绸,2026,63(1):57-67,11.基金项目
陕西省服务地方专项重点项目(14JF008) （14JF008）
