Open-Ended Evolution of Artistic Styles in Diffusion Models via Island-Based Genetic Algorithms

This paper explores a novel approach to generative image creation by evolving prompt embeddings—the internal conditioning tensors used by diffusion models—rather than optimizing prompt text. Using SDXL-Turbo and a real-valued Genetic Algorithm within an Island Model setup, distinct artistic styles are assigned to separate islands, which evolve independently and interact via migration to stimulate stylistic influence.