OPTIMIZING GENERATIVE AI BY OVERCOMING STABILITY MODE COLLAPSE AND QUALITY CHALLENGES IN GANS AND VAES

Abstract

This paper reviews the current state of Generative AI, focusing on two of the most prominent models: GANs, an abbreviation for Generative Adversarial Networks and VAEs, an abbreviation for Variational Autoencoder. Despite the great potential of these models in producing high quality synthetic data, there are some issues that have been observed which include stability, mode collapse and quality of the generated results. There are many problems that arise during GAN training, including instability, which causes such problems as mode collapse, when the generator generates a small number of images or similar images. While running through the network, VAEs are observed to be more stable and less noisy than the GANs but they are not as accurate as the GANs in generation of samples. The mentioned limitations have been addressed by recent developments including Wasserstein GANs, feature matching, progressive GANs, and a combination of both such as VAE-GANs. These innovations are intended to increase stability and sample quality to employ new loss functions, training methods, as well as the new architectures that incorporate advantages of GANs and VAEs. However, several issues are still unresolved, these are computational cost, growth, and the question of the potential malicious use of the generative AI tools. This paper also discusses potential future research topics in the field, including self-supervised learning, combined multimodal methods, and the introduction of ethical measures when implementing generative models. The purpose of this review is to present the current state of the art, as well as current and potential issues in generative models, and different strategies to further increase the performance of these algorithms.