Diffusion models (panel (a)) link a data-generating distribution (high-resolution DEMs) to the Gaussian distribution (white noise) via an SDE, while SBs (panel (b)) link two arbitrarily complicated distributions. The reverse-time dynamics of these samplers can be modeled using vision transformers (panel (c)). Representing these models in the latent space of a variational autoencoder (VAE; panel (d)) is often more efficient.