Diffusion Models in Vision: A Survey.

Area of Science:

• Computer Vision
• Deep Generative Modeling

Background:

• Denoising diffusion models are emerging deep generative models.
• They involve a forward process of adding noise and a reverse process of noise removal.
• These models achieve high-quality and diverse sample generation but face computational challenges.

Purpose of the Study:

• To provide a comprehensive review of denoising diffusion models in computer vision.
• To categorize existing diffusion model frameworks and discuss their relationships with other generative models.
• To identify current limitations and suggest future research avenues.

Main Methods:

• Review of theoretical and practical contributions in denoising diffusion models.
• Identification of three generic diffusion modeling frameworks: denoising diffusion probabilistic models, noise conditioned score networks, and stochastic differential equations.
• Discussion of the relationships between diffusion models and other deep generative models (VAEs, GANs, etc.).

Main Results:

• Categorization of diffusion models in computer vision from multiple perspectives.
• Analysis of the strengths and weaknesses of different diffusion modeling frameworks.
• Comparison of diffusion models with established generative models like VAEs and GANs.

Conclusions:

• Denoising diffusion models are powerful tools for generative tasks in computer vision.
• Understanding their theoretical underpinnings and practical applications is crucial.
• Further research is needed to address computational burdens and explore new frontiers.