Residual-Space Evolutionary Optimization via Flow-based Generative Models

The article presents a novel framework called residual-space evolutionary optimization, which integrates flow-based generative models with evolutionary algorithms to enhance data editing capabilities. This model-agnostic approach leverages conditional flow matching (CFM) to operate in residual space, allowing for targeted local exploitation and broader exploration of data. Validation on the MorphoMNIST dataset and crystal data showcases its effectiveness in balancing target alignment, instance preservation, and diversity, making it relevant for practitioners aiming to improve generative editing in both synthetic and scientific applications.