Deep Generative Models for Image Synthesis: GANs, VAEs, and Diffusion Approaches

Abstract

This paper provides a comprehensive overview of the evolution and current state of deep generative models in image synthesis, with a focus on Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and diffusion models. It identifies persistent challenges, including difficulty in generating high-resolution details, training instabilities, and limited controllability. The analysis demonstrates that Diffusion Models have emerged as the dominant paradigm, effectively overcoming previous limitations through their robust training dynamics and superior detail preservation. By modelling data distribution via gradual noise addition and removal, diffusion models achieve state-of-the-art image quality while maintaining structural coherence. Their integration with text-conditioning mechanisms enables precise semantic control, facilitating diverse applications such as high-fidelity text-to-image synthesis, medical image enhancement, and accelerating creative design. The review further examines practical implementations, including SRGAN for super-resolution, CVAE for image restoration, and diffusion-based text-to-image systems, highlighting their real-world impact across healthcare, media, and scientific imaging. The work concludes that diffusion models, particularly with latent-space optimisation, represent the most promising direction for future research, balancing unprecedented quality with increasing computational efficiency to enable broader deployment in critical domains.