Iguchi & Fearnhead prove diffusion models adapt to intrinsic data dimension

Diffusion models have become the backbone of generative AI, but their theoretical underpinnings lag behind practice—especially for real-world data that isn't a single smooth manifold. Iguchi and Fearnhead tackle multimodal, high-dimensional data: multiple clusters in R^D, each with its own low-dimensional structure. They model this with K-mixture Gaussians and prove two results. First, the denoising process can be interpreted as a dynamical Bayesian classifier: the score function is a posterior-weighted average of cluster-wise scores. With high probability, posterior class probabilities concentrate on a single cluster once the signal-to-noise ratio reaches scale Θ(log(KD)/D). This matches when the model "commits" to a specific cluster during generation.

Second, by splitting the denoising into mixing and cluster-commitment phases, they derive a KL error bound that depends linearly on the maximum intrinsic dimension of a cluster, up to a logarithmic factor—even as K grows polynomially with D. This is a direct improvement over ambient-dimensional bounds and generalizes earlier low-dimensional adaptivity results to multimodal distributions with heterogeneous, approximately low-rank covariances. Practically, this suggests diffusion models automatically exploit low-dimensional structure without needing to know it in advance, which could inspire more efficient training and sampling strategies. The paper also provides a clean theoretical framework for when class information emerges during generation, relevant for conditional generation and inverse problems.