Can Diffusion Models Disentangle? A Theoretical Perspective

Liming Wang, Yuan Gong, Marco Vilela, James R. Glass

Keywords: diffusion model, disentanglement, voice conversion

TL;DR: A theory of the disentanglement ability of diffusion models

Abstract: This paper introduces a novel theoretical framework to understand how diffusion models can learn disentangled representations under the assumption of an $\normltwo$ score approximation. We also provide sufficient conditions under which such representations are beneficial for domain adaptation. Our theory offers new insights into how existing diffusion models disentangle latent variables across general distributions and suggests strategies to enhance their disentanglement capabilities. To validate our theory, we perform experiments using both synthetic data generated from latent subspace models and real speech data for non-parallel voice conversion - a canonical disentanglement problem. Across various classification tasks, we found voice conversion-based adaptation methods achieve significant improvements in classification accuracy, demonstrating their effectiveness as domain adaptors. Code will be released upon acceptance.

Primary Area: learning theory

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: 7744