How Generative Models Approach Molecular Conformational Sampling

Characterising equilibrium conformational ensembles with deep generative models requires assessing not only whether a model reproduces the target distribution, but also the mechanism of how it arrives here. Here we examine two distinct routes to generative conformational sampling-- stochastic relaxation and deterministic transport--through a study of denoising diffusion probabilistic models (DDPM) and rectified-flow (RF) models across molecular systems of increasing complexity. Using systems of increasing complexity, including a multimodal two-dimensional potential, the folded mini-protein Trp-cage, and a high-dimensional dihedral subspace of the intrinsically disordered protein -synuclein, we show that the key distinction between these paradigms lies not only in endpoint fidelity but in how distributional error is resolved during sampling. Diffusion models converge via pronounced late-stage stochastic relaxation and exhibits robust recovery of configurational breadth across neural architectures. Rectified flow approaches the target more gradually through deterministic transport and therefore depends much more strongly on architectural expressivity, particularly in heterogeneous high-dimensional landscapes. Analyses of entropy and moment evolution further show that diffusion more reliably restores both ensemble location and fluctuation structure, whereas RF requires Transformer-level feature mixing to represent the transport geometry accurately. These results establish convergence mechanism as a key design principle for generative sampling.