Learning goal: these samplers use the same score model and starting samples while changing η, the stochasticity setting. Under this demo's update convention, η=1 is stochastic and η=0 deterministic.

Pale orange: data; thin: sampling paths; bright: endpoints. The score is computed from the empirical Gaussian-smoothed data distribution, without neural approximation error.

Reduce η from 1 to 0: paths become less stochastic and more deterministic. Exact continuous-time marginal identities do not guarantee identical finite-step distributions. Reduce steps to inspect discretization effects; nearest-point distance is only one diagnostic. This empirical exact-score toy does not test neural score-estimation error or prove a universal sampler ranking.