Derive Evolutionary Methods from Alternative Diffusion Model Implementations

Investigate whether alternative diffusion model implementations can be leveraged to construct evolutionary algorithms that exhibit distinct properties and diversity characteristics relative to Diffusion Evolution, and determine the specific features and behaviors these variants introduce.

Background

The authors argue that diffusion models and evolution are two sides of the same process and demonstrate Diffusion Evolution using DDIM-style sampling with particular schedules.

They suggest that variations in diffusion model design may yield different evolutionary dynamics, potentially enabling unique features; exploring this would clarify the scope of the diffusion–evolution equivalence and its practical consequences.