Evolutionary Optimization in Residual Space for Generative Data Editing

Data editing using generative methods typically requires differentiable objectives and gradient-based search techniques. However, these assumptions often break down in flow-based settings, where editing involves forward and backward integration and objectives can be non-differentiable or black-box. This research addresses this challenge by proposing residual-space evolutionary optimization, a framework that is agnostic to the specific model used. The framework integrates flow-based generative editing with evolutionary algorithms. It builds on the insight that conditional flow matching (CFM) can effectively disentangle factors controlled by conditions from instance-specific residuals. By directly operating in this residual space, the framework separates two complementary search strategies: "self-pollination" for local exploitation through refining feature-preserving residuals, and "cross-pollination" for broader exploration by recombining residuals across diverse samples. As a proof of concept, the method was validated on MorphoMNIST, a benchmark for counterfactual generation, and on crystal data. The results demonstrate that this decomposition into exploration and exploitation provides an effective mechanism for balancing target alignment, instance preservation, and diversity. The approach extends beyond image data to real-world scientific domains, offering a versatile tool for generative data editing.