Sampling from the terminal distribution of a stochastic differential equation (SDE) is a core step in diffusion-based generative AI. The standard approach draws independent sample paths, but a fixed simulation budget limits accuracy. Gupta and Juneja propose instead splitting partial paths into a tree, reusing computation across correlated branches.

The worry with such splitting is that the induced dependence could bias the empirical distribution. The authors analyze this directly, constructing a general splitting design and deriving limiting errors for the Kolmogorov-Smirnov distance and maximum mean discrepancy as the budget grows. They also characterize a splitting strategy based on the associated asymptotic optimization problem.

In experiments, the method improved mean KS by 12–26% over i.i.d. sampling across many settings, and on CIFAR-10 and FFHQ it lowered MMD by 17–26%. The results suggest that carefully designed dependence can be a practical advantage rather than a liability in SDE-based sampling.