Image Synthesis Using GANs and Diffusion Models

Abstract

Image synthesis is the creation of new images using existing ones as a reference or source. This can be done using various techniques, such as image manipulation, computer graphics, and deep learning. In Deep Learning, image synthesis can be achieved using Generative Adversarial Networks (GANs), Diffusion Models (DMs), Variational Autoencoders (VAEs), etc. GANs and Diffusion Models are becoming a hot topic in generative models and image synthesis. In this paper, we put forth the recent research on GANs and Diffusion Models for image synthesis. GANs and Diffusion Models have successfully generated high-quality images in various domains, such as faces, animals, landscapes, etc. We then discuss these models' advantages, disadvantages, and unique features. We get familiar with the general architecture of DMs and GANs for image synthesis and how they can generate realistic-looking images with very high caption similarity. We also touch on Inception Score (IS) and Frechet Inception Distance (FID) and how they are used for ' various image synthesis methods as evaluation criteria. We briefly explain why IS and FID should not be the only performance criteria a model should adopt and why human evaluation is also essential for the generated image. This review aims to provide insights into the research on deep learning techniques such as GANs and Diffusion Models.