Unified Perspectives on Signal-to-Noise Diffusion Models

Khanh Doan, Trung Le, Long Tung Vuong, Anh Tuan Bui, Quyen Tran, Thanh-Toan Do, Mehrtash Harandi, Dinh Phung

Keywords: diffusion models, SDE, ODE, continuous variation models

Abstract: Diffusion models (DM) have become essential components of generative modeling, demonstrating exceptional performance in domains like image synthesis, audio generation, and complex data interpolation. Signal-to-Noise diffusion models represent a broad family encompassing many state-of-the-art models. Although several efforts have been made to explore Signal-to-Noise (S2N) diffusion models from different angles, a comprehensive study that connects these viewpoints and introduces new insights is still needed. In this work, we provide an in-depth perspective on noise schedulers, analyzing their role through the lens of the signal-to-noise ratio (SNR) and its relationship to information theory. Based on this framework, we introduce a generalized backward equation to improve the efficiency of the inference process.

Primary Area: generative models

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: 8076