基于扩散模型的两阶段服装图像生成方法研究  

Research on the Two-stageclothing Image Generation Method Based on a Diffusion Model

作　　者：叶青明 徐亿波 刘正[3,4,5] 杨阳[1,3] 侯珏 YE Qingming;XU Yibo;LIU Zheng;YANG Yang;HOU Jue(School of Fashion Design&Engineering,Hangzhou,Zhejiang 310018,China;Excellent Fashion Garment(Hangzhou)Co.,Ltd,Hangzhou,Zhejiang 310018,China;Zhejiang Provincial Engineering Laboratory of Clothing Digital Technology,Hangzhou,Zhejiang 310018,China;International Institute of Fashion Technology,Hangzhou,Zhejiang 310018,China;Key Laboratory of Silk Culture Heritage and Products Design Digital Technology,Ministry of Culture and Tourism,Zhejiang Sci-Tech University,Hangzhou,Zhejiang 310018,China)

机构地区：[1]浙江理工大学服装学院,浙江杭州310018 [2]卓尚服饰(杭州)有限公司,浙江杭州310018 [3]浙江理工大学服装数字化技术浙江省工程实验室,浙江杭州310018 [4]浙江理工大学国际时装技术学院,浙江杭州310018 [5]浙江理工大学丝绸文化传承与产品设计数字化技术文化和旅游部重点实验室,浙江杭州310018

出　　处：《北京服装学院学报(自然科学版)》2025年第1期86-93,共8页Journal of Beijing Institute of Fashion Technology:Natural Science Edition

基　　金：中华人民共和国文化和旅游部重点实验室资助项目《云锦数字化解构与智能设计》(23072247-N)。

摘　　要：文生图问题是人工智能技术中的一个重要分支,常用于服装设计、纹样绘制等场景。然而,现有算法的直接文生图方法难以对生成服装图像的款式、颜色花纹等信息进行精准控制。为此,本文提出了一种基于扩散模型的两阶段图像生成方法,将服装图像的生成任务分离为款式-颜色纹理2个阶段,使模型能够更加准确捕捉文本信息中的款式信息与颜色信息。在第1阶段中,利用LoRA微调Stable Diffusion模型,基于文本信息生成精确的边缘线稿草图以表达服装款式信息;第2阶段结合ControlNet模型,将生成的线稿草图与文本中的颜色花纹信息融合,生成最终图像。同时,本研究设计了款式颜色信息过滤模型,将款式和颜色信息与普通信息分离,并赋予两者较大权重,从而增强模型对有效信息的捕捉能力。为了验证两阶段方法的有效性,通过FID(弗雷谢起始距离)、PSNR(结构相似指数)和SSIM(峰值信噪比)等指标进行客观评价。结果表明,采用该两阶段方法生成的服装图像在款式、颜色花纹和细节处理上显著优于其他方法。The text-to-image generation problem is a significant branch of artificial intelligence technology,widely used in scenarios such as fashion design and pattern drawing.However,existing direct text-to-image methods face challenges in precisely controlling the style,color,and pattern information in generated clothing images.To address this,this paper proposes a two-stage image generation method based on a diffusion model,separating the clothing image generation task into two phases:text-style and color-texture,enabling the model to more accurately capture style and color information from the text.In the first stage,a LoRA fine-tuned Stable Diffusion model is used to generate precise edge line sketches based on the text,expressing the style information of the clothing.In the second stage,the ControlNet model is combined to integrate the generated sketches with the color and pattern information from the text,producing the final image.Additionally,a style-color information filtering model is designed to separate style and color information from general information,assigning greater weight to both,thereby enhancing the model’s ability to capture relevant information.To validate the effectiveness of the two-stage method,objective evaluations were conducted using metrics such as FID,PSNR,and SSIM.The results demonstrate that the clothing images generated by this two-stage approach significantly outperform other methods in terms of style,color patterns,and detail handling.

关 键 词：服装图像生成 扩散模型 CONTROLNET网络 虚拟试衣 服装设计 

分 类 号：TS941.26[轻工技术与工程—服装设计与工程]