Making images with AI often uses a process where noise is gradually added to data, and a neural network learns to reverse this. Reversing it randomly (stochastic sampling) gives richer images than the strict path (deterministic sampling) but takes many more steps, making it slow to use.
Common tricks to speed up the strict path fail for the random path because the noise breaks the higher-order math. The authors introduce forward-value discretization, which evaluates the neural network at a predicted future state (a lookahead) requiring only one network call per step.
Their new stochastic sampler produces better images (lower Fréchet Inception Distance) than existing stochastic samplers at the same budget, and competitive quality with the strict path using only 1.5 times the steps. This reduces the practical efficiency gap between the two approaches.