Residual-Space Evolutionary Optimization via Flow-based Generative Models
Researchers have developed a new framework called residual-space evolutionary optimization, designed to enable data editing with generative models, particularly in flow-based settings where traditional gradient-based methods are not applicable. This model-agnostic approach combines flow-based generative editing with evolutionary algorithms, operating directly in the residual space to separate local exploitation and broader exploration. The framework has been demonstrated on the MorphoMNIST dataset for counterfactual generation and on crystal data, showing its effectiveness in balancing target alignment, instance preservation, and diversity across different domains. AI
IMPACT This new framework could enable more flexible and powerful data editing capabilities for generative AI models, particularly in scientific applications.