A Comparative Analysis of Diffusion-Based Models in Text-to-Image Generation

Abstract

The recent years have seen the rapid development of artificial intelligence image generation methods with the introduction of a wide range of different models to produce images based on a specific text description. The paper explores the design, functionality, and restrictions of four well-known diffusion-based text-to-image generative artificial intelligence systems: Midjourney, DALL-E2, Stable Diffusion, and Imagen. The paper defines the theoretical context of the diffusion probabilistic model and the manner in which it has been applied in the four models of artificial intelligence. This paper uses the comparative and evaluation performance of these models based on the available empirical research studies on the same in relation to image fidelity, prompt adherence, creativity, and bias on various parameters. The results of the analysis show that Stable Diffusion works well in the generation of photorealistic pictures of the human face, Midjourney works well in creativity and visual image in artistic settings, and Imagen works well in complex text description. These models have also exhibited several challenges and limitations. The findings depict the difficulties in the production of precise, just, and imaginative imagery in a variety of situations.