Disentangling total-variance and signal-to-noise-ratio improves diffusion models
Machine Learning: Science and Technology · 01.12.2026 · 7 (6) · 065002
DOI: 10.1088/2632-2153/aeb69d ↗Kurz gefasst
Fachartikel in der Zeitschrift Machine Learning: Science and Technology, online erschienen am 30.10.2026. Verfasst von Khaled Kahouli und 5 weiteren Autor:innen. Die Arbeit nennt 74 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: German Ministry for Education and Research, Korean government, Institute of Information & communications Technology Planning & Evaluation.
Abstract (Originalsprache)
Diffusion models have been widely adopted as a powerful generative framework across numerous scientific domains, yet their inherently iterative sampling process remains a significant computational bottleneck. While this is typically mitigated by reducing the number of reverse diffusion steps, sample quality becomes highly sensitive to the chosen noise schedule, i.e. the manner in which noise is introduced and the signal is reduced at each step. Although prior work has improved on the original variance-preserving and variance-exploding schedules, these approaches passively adjust the total variance, without direct independent control over it. In this work, we propose a novel total-variance/signal-to-noise-ratio disentangled framework, where total variance (TV) and signal-to-noise-ratio (SNR) can be controlled independently. Our approach reveals that schedules with exponentially exploding TV can be consistently improved by adopting a constant TV schedule while preserving the underlying SNR schedule. Furthermore, we demonstrate that generalizing the SNR schedule of optimal transport flow matching significantly enhances generation efficiency. Crucially, these gains remain consistent across various reverse diffusion solvers and reduce the search space to two independent and interpretable parameters, enabling fast and robust sampling for complex scientific applications, such as molecular structure generation.
Förderung
- German Ministry for Education and Research
- Korean government
- Institute of Information & communications Technology Planning & Evaluation
- Swiss National Science Foundation
- Hector Fellow Academy
Zitieren
APA
Kahouli, K., Ripken, W., Gugler, S., Unke, O. T., Müller, K., & Nakajima, S. (2026). Disentangling total-variance and signal-to-noise-ratio improves diffusion models. Machine Learning: Science and Technology, 7(6), 065002. https://doi.org/10.1088/2632-2153/aeb69d
Vancouver
Kahouli K, Ripken W, Gugler S, Unke OT, Müller K, Nakajima S. Disentangling total-variance and signal-to-noise-ratio improves diffusion models. Machine Learning: Science and Technology. 2026;7(6):065002. doi:10.1088/2632-2153/aeb69d
BibTeX
@article{Kahouli2026,
author = {Kahouli, Khaled and Ripken, Winfried and Gugler, Stefan and Unke, Oliver T and Müller, Klaus-Robert and Nakajima, Shinichi},
title = {Disentangling total-variance and signal-to-noise-ratio improves diffusion models},
journal = {Machine Learning: Science and Technology},
year = {2026},
volume = {7},
number = {6},
pages = {065002},
doi = {10.1088/2632-2153/aeb69d}
}
EUVIMED