Evolution, Comparison, and Trend Prospects of Text Generated Images Based on Deep Learning

Abstract

Text generated image is a hot research field of cross modal in-depth learning, which can effectively replace the traditional manual drawing and be applied in games, advertising and other industries. The existing mainstream technologies are divided into Generative Adversarial Network (GAN), transformer and diffusion model. Each model has undergone multiple rounds of structural iterative optimization. This paper sorts out the evolution of the three types of architectures, and integrates the open quantitative indicators such as Fréchet Inception Distance (FID), Inception Score (IS), and image text matching of various models under the COCO dataset to compare the performance differences of existing schemes. GAN series models have fast reasoning speed but low image quality limit, transformer has strong detail expressiveness but huge computational cost, and diffusion model has real-time imaging but insufficient real-time performance. At present, there are many pain points in this field, such as the difficulty of taking into account the computational power, image quality and speed at the same time, the deviation of long text semantic matching, and the weak ability of scene generalization. Based on this, this paper summarizes the mainstream optimization research directions, such as lightweight, multi architecture integration, and multi condition controllable generation.