[1]GUAN Fengxu,ZHANG Hanyu,LU Siqi,et al.Research status of diffusion models in computer vision[J].CAAI Transactions on Intelligent Systems,2025,20(2):265-282.[doi:10.11992/tis.202312041]
Copy
CAAI Transactions on Intelligent Systems[ISSN 1673-4785/CN 23-1538/TP] Volume:
20
Number of periods:
2025 2
Page number:
265-282
Column:
综述
Public date:
2025-03-05
- Title:
-
Research status of diffusion models in computer vision
- Author(s):
-
GUAN Fengxu; ZHANG Hanyu; LU Siqi; LAI Haitao; DU Xue; ZHENG Yan
-
College of Intelligent Systems Science and Engineering, Harbin Engineering University, Harbin 150001, China
-
- Keywords:
-
diffusion model; denoising diffusion probabilistic model; score-based generative model; deep learning; computer vision; image generation; generative model; generative adversarial network
- CLC:
-
TP18
- DOI:
-
10.11992/tis.202312041
- Abstract:
-
The diffusion model is a new generative model inspired by molecular thermodynamics. This model offers stable training and low dependence on model settings, making it a popular benchmark in computer vision. In recent years, the diffusion model has been widely applied to various tasks, yielding diverse and high-quality results compared to traditional generative models. At present, the diffusion model is a prominent method in the field of computer vision. This paper provides a comprehensive overview of the diffusion model to further stimulate its development in this domain. First, the paper compares the advantages and disadvantages of diffusion models with other generative models and introduces the underlying mathematical principles. Then, the study presents recent efforts by researchers to improve diffusion models, starting with common challenges and highlighting application examples in various visual tasks. Finally, the study discusses existing issues with diffusion models and outlines potential future development trends.