Deep generative models are the key driving force behind the breakthroughs in image synthesis in the field of artificial intelligence, aiming to learn complex data distributions and sample from them to generate realistic and diverse new images. This article aims to systematically trace the evolution process of this field from its pioneering models to the current cutting-edge technologies. The paper clearly presents this development process, the core review logic of this paper is to divide the mainstream deep generative models into three major frameworks based on the core modeling ideas and learning paradigms of the models, namely Generative Adversarial Network (GAN), Variational AutoEncoder (VAE), and Diffusion Model (DM). Based on this logic, this paper first expounds the basic principles of various frameworks, representative derivation methods and the inheritance relationships among them; Furthermore, the system compared the generation performance of different models on multiple benchmark datasets; Finally, the challenges currently faced by deep generation technology were deeply discussed, and its future research directions were prospected. Through the review, this article hopes to provide researchers with a clear panoramic view of technological development and inspire new research ideas.
Zhiming Hou (Mon,) studied this question.