Adaptive Second-Order Solvers for Fast Stochastic Diffusion Sampling

Adaptive PI step-size control stabilizes diffusion sampling by incorporating previous error estimates into the current step adjustment. This noise-normalized error estimation enables smoother step adaptation across varying trajectories. Aggregating these per-sample trajectories into fixed schedules maintains most efficiency gains for natural-image and language datasets.