ImageBreeder : guiding diffusion models with evolutionary computation

Abstract

With the recent advancements of diffusion models, it is quite easy to generate high-quality images. However, many attempts and manual changes are often necessary to achieve this high quality. Leveraging evolutionary algorithms to automate this process therefore presents a promising approach. Consequently, we introduce Image-Breeder as a framework to improve image generation at inference time driven by evolutionary algorithms. Additionally, we study the effectiveness of 10 different variation operators ranging from pixel-based blending techniques to modifications directly on the latent representation of the images. The results show that using evolutionary algorithms can significantly increase an image's quality as well as the alignment with a given prompt. Furthermore, all tested guided search methods are human competitive, as a random trial and error approach is outperformed on over 75% of the benchmark problems. In addition to the empirical results, we also critically discuss the benefits and challenges of our framework uncovering future research directions. We recommend researchers to focus on optimising latent image representations for this purpose as this may improve the ability of evolutionary algorithms to transfer promising features from one image to another, unlocking even better image generation.