Diffusion Models Adapt to Low-Dimensional Structure Under Flexible Coefficient Choices

arXiv·low signal

Cai, Jiao, and Li develop convergence theory showing diffusion models exploit unknown low-dimensional structure to accelerate sampling even under flexible coefficient (noise-schedule) choices, generalizing prior results that assumed narrower settings. The analysis helps explain why diffusion samplers converge faster than worst-case bounds suggest. Primarily theoretical but informs sampler and schedule design for generative-model practitioners.