Words shaping worlds: A comprehensive exploration of text-driven image and video generation with generative adversarial networks

Abstract

A growing interest in education, media, and entertainment has increased Artificial Intelligence-powered content generation, mainly via Generative Adversarial Networks (GANs), which shape images, videos, audio, and text. A generative adversarial network (GAN) is the combination of two deep neural networks: generator ($G$) and discriminator ($D$). These two components are trained competitively by pitting one against the other such that $G$ generates new data while $D$ authenticates the data. By leveraging powerful deep neural networks and competitive training, GANs can synthesize reasonable and realistic images and videos from the text description. This paper extensively reviews the recent state-of-the-art GAN models for text-to-image (T2I) and text-to-video (T2V) synthesis. In this regard, databases like ACM, IEEE Explore, Web of Science, and ScienceDirect were searched to find and analyze the relevant research articles conducted in this area in the last decade, specifically from 2014 to 2024. Secondly, T2I and T2V GAN methods were classified according to structure and functionality. Later, a comprehensive evaluation between T2I and T2V GAN-based methods was conducted, employing various qualitative and quantitative evaluation techniques. Finally, the paper concludes by discussing multiple applications, main challenges, and limitations of T2I and T2V GAN models for future consideration.