Use Evolutionary Insights to Enhance Diffusion Models

Determine which insights from evolutionary theory—such as selection, mutation, reproductive isolation, open-endedness, and novelty search—can be systematically incorporated into diffusion model training or sampling procedures to improve generative performance, diversity, and controllability.

Background

The paper presents a bidirectional connection: diffusion models can perform evolutionary algorithms, and evolutionary principles can inform generative modeling.

Formalizing how evolutionary mechanisms translate into modifications of diffusion training or sampling could yield new classes of generative models with improved robustness and diversity, leveraging the synergy the authors describe.