Diffusion Models in Analog Circuits

Physics-based diffusion models—the foundation of today’s generative AI— are implemented abstractly with digital central processing, resulting in enormous computational and energy costs. We take the opposite approach: a physical learning system whose observables directly encode the statistics of a diffusion process. We develop an algorithm that enables this system to infer the reverse dynamics solely from forward trajectories and apply it to the problem of image generation. The hardware comprises a network of self-adjusting resistors governed by the coupled local learning framework, exploiting the mathematical equivalence between voltages in passive networks and absorption probabilities in Markov chains. This architecture supports fully parallel operation and can be realized in scalable, ultra-low-power hardware. We present ongoing work toward a microfabricated implementation using transistors biased in their subthreshold regime.