Efficient Diffusion Models for Symmetric Manifolds

Oren Mangoubi, Nisheeth K. Vishnoi

Keywords: Diffusion model, symmetric manifolds

Abstract:

We present a framework for designing efficient diffusion models on symmetric Riemannian manifolds, which include the torus, sphere, special orthogonal group, and unitary group. While diffusion models on symmetric manifolds have gained significant attention, existing approaches often rely on the manifolds' heat kernels, which lack closed-form expressions and result in exponential-in-dimension per-iteration runtimes during training. We introduce a new diffusion model for symmetric-space manifolds, leveraging a projection of Euclidean Brownian motion to bypass explicit heat kernel computations. Our training algorithm minimizes a novel objective function derived via Ito's Lemma, with efficiently computable gradients, allowing each iteration to run in polynomial time for symmetric manifolds. Additionally, the symmetries of the manifold ensure the diffusion satisfies an "average-case" Lipschitz condition, enabling accurate and efficient sample generation. These improvements enhance both the training runtime and sample accuracy for key cases of symmetric manifolds, helping to bridge the gap between diffusion models on symmetric manifolds and Euclidean space.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4845