摘要
扩散模型在生成模型领域具有高质量的样本生成能力,一经推出就不断地刷新图像生成评价指标FID分数的记录,成为了该领域的研究热点,而此类相关综述在国内还鲜有介绍。因此,文中对相关扩散生成模型的研究进行汇总与分析。首先,对去噪扩散概率模型、基于分数的扩散生成模型和随机微分方程的扩散生成模型这3类通用模型的特点和原理进行了论述,就每一类基本扩散模型中以优化模型内部算法、高效采样为改进目标的相关衍生模型进行分析。其次,对当下扩散模型在计算机视觉、自然语言处理、时间序列、多模态和跨学科领域等方面的应用进行总结。最后,基于上述论述,分别就目前扩散生成模型存在的采样步骤多、采样时间长等局限性提出了相关建议,并结合前述研究对未来扩散生成模型的发展方向进行了研判。
Diffusion models have shown high-quality sample generation ability in the field of generative models,and constantly set new records for image generation evaluation indicator FID scores since their introduction,and has become a research hotspot in this field.However,related reviews of this kind are scarce in China.Therefore,this paper aims to summarize and analyze the research on related diffusion generative models.Firstly,it analyzes the related derivative models in each basic diffusion model,which focus on optimizing internal algorithms and efficient sampling,by discussing the characteristics and principles of three common models:denoising diffusion probabilistic model,score-based diffusion generative model,and diffusion generative model based on random differential equations.Secondly,it summarizes the current applications of diffusion models in computer vision,natural language processing,time series,multimodal,and interdisciplinary fields.Finally,based on the above discussion,relevant suggestions for the existing limitations of diffusion generative models are proposed,such as long sampling times and multiple sampling steps,and a research direction for the future development of diffusion generative models is provided based on previous studies.
作者
闫志浩
周长兵
李小翠
YAN Zhihao;ZHOU Zhangbing;LI Xiaocui(School of Information Engineering,China University of Geosciences(Beijing),Beijing 100083,China)
出处
《计算机科学》
CSCD
北大核心
2024年第1期273-283,共11页
Computer Science
基金
国家自然科学基金(42050103)。
关键词
深度学习
生成模型
去噪扩散概率模型
基于分数的扩散模型
随机微分方程
图像生成
Deep learning
Generative models
Denoising diffusion probabilistic models
Score-based diffusion models
Stochastic differential equations
Image generation
