基于扩散模型的图像生成方法研究综述  

A review of image generation methods based on diffusion models

作　　者：龚帅 邓勇 向金海 GONG Shuai;DENG Yong;XIANG Jinhai(Key Laboratory of Smart Farming Technology for Agricultural Animals,Ministry of Agriculture and Rural Affairs,Wuhan 430070,China;Engineering Research Center of Intelligent Technology for Agriculture,Ministry of Education,Wuhan 430070,China;School of Information,Huazhong Agricultural University,Wuhan 430070,China;722 Research Institute of China Shipbuilding Co.,Ltd.,Wuhan 430205,China)

机构地区：[1]农业农村部智慧养殖技术重点实验室,湖北武汉430070 [2]农业智能技术教育部工程研究中心,湖北武汉430070 [3]华中农业大学信息学院,湖北武汉430070 [4]中国船舶集团有限公司第七二二研究所,湖北武汉430205

出　　处：《武汉大学学报(工学版)》2025年第2期292-305,共14页Engineering Journal of Wuhan University

摘　　要：图像生成任务是计算机视觉中的一个重要研究领域,基于深度学习的相关工作日益增多。扩散模型作为深度学习中的一类新兴的生成模型,发展迅速,大大促进了图像生成算法的发展,因此对基于扩散模型的图像生成方法系统地开展文献综述非常有必要。从加速采样、可控性和图像编辑3个方面对基于扩散模型的图像生成方法进行大量分析和研究。首先,对比分析了基于概率和基于分数匹配的图像生成算法的局限性,并介绍了对应加速采样算法的发展;随后,通过对基于扩散模型的图像生成方法中可控性方面的研究进行综述,对比指出相关方法的优势和应用场景;此外,讨论了不同的图像编辑方法,分析比较了每种模型的优点及不足;最后,指出了基于扩散模型的图像生成方法值得进一步研究的问题和发展方向。Image generation is an important research field in computer vision,and the related work based on deep learning is increasing.As an emerging generative model in deep learning,diffusion model is developing rapidly,which greatly promotes the development of image generation algorithms.Therefore,it is necessary to carry out a systematic literature review of image generation methods based on diffusion model.A lot of analysis and research on image generation methods based on diffusion model are carried out in terms of accelerating sampling,controllability and image editing.Firstly,the limitations of image generation algorithms based on probability and score-matching are compared and analyzed,and the development of corresponding accelerating sampling algorithm is introduced.Then,by reviewing the research on controllability of image generation methods based on diffusion model,the advantages and application scenarios of related methods are compared.Besides,different image editing methods are discussed,and the advantages and disadvantages are analyzed and compared.Finally,the problems and the development directions of image generation based on diffusion model worthy of further study are pointed out.

关 键 词：人工智能 深度学习 扩散模型 图像生成 图像编辑 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]