From GANs to Diffusion: The Evolution of Text-to-Image Generation Techniques

Abstract

Text-to-Image (T2I) technique has become one of the most important research topics of Artificial Intelligence (AI). It mainly focuses on generating high-quality images through natural language. In the past 10 years, this field developed rapidly: 1) GAN-based approaches; 2) Transformer-based models; 3) diffusion models. Every technique paradigm improves the disadvantages of former one and develops new strengths, such as image quality, semantic alignment, stability, and scalability. The paper systematically explores the evolution of T2I technique. First introduces three main kinds of models: GAN-based models, Transformer-based models, and diffusion models. Following that are the key techniques of each phase including text representation, cross-model alignment, generative mechanisms, and efficiency optimization. The challenges which remain to be resolved will be discussed in the next part. In the end, future application and research directions will be displayed. Overall, this paper aims to show the major evolution and process of the development of T2I techniques. It also offers insights into its potential development and commercial applications.