Abstract
Score-based diffusion models generate samples from a complex underlying datadistribution by time-reversal of a diffusion process and represent thestate-of-the-art in many generative AI applications. Here, I show how agenerative diffusion model can be implemented based on an underlying activeparticle process with finite correlation time. Time reversal is achieved byimposing an effective time-dependent potential on the position coordinate,which can be readily implemented in simulations and experiments to generate newsynthetic data samples driven by active fluctuations. The effective potentialis valid to first order in the persistence time and leads to a force field thatis fully determined by the standard score function and its derivatives up to2nd order. Numerical experiments for artificial data distributions confirm thevalidity of the effective potential, which opens up new avenues to exploitfluctuations in active and living systems for generative AI purposes.