![[2602.19512] Variational Trajectory Optimization of Anisotropic Diffusion Schedules](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.19512] Variational Trajectory Optimization of Anisotropic Diffusion Schedules
Summary
This paper presents a variational framework for optimizing anisotropic diffusion schedules in machine learning, enhancing performance across various datasets.
Why It Matters
The study addresses a critical aspect of diffusion models by proposing a novel optimization technique that improves model efficiency and accuracy. This advancement is significant for researchers and practitioners in machine learning and computer vision, as it can lead to better performance in generative tasks.
Key Takeaways
- Introduces a variational framework for anisotropic noise schedules.
- Implements a trajectory-level objective for joint training of score networks.
- Develops an efficient reverse-ODE solver for inference.
- Demonstrates consistent performance improvements over baseline models on multiple datasets.
- Provides accessible code for implementation and further research.
Computer Science > Machine Learning arXiv:2602.19512 (cs) [Submitted on 23 Feb 2026] Title:Variational Trajectory Optimization of Anisotropic Diffusion Schedules Authors:Pengxi Liu, Zeyu Michael Li, Xiang Cheng View a PDF of the paper titled Variational Trajectory Optimization of Anisotropic Diffusion Schedules, by Pengxi Liu and 2 other authors View PDF HTML (experimental) Abstract:We introduce a variational framework for diffusion models with anisotropic noise schedules parameterized by a matrix-valued path $M_t(\theta)$ that allocates noise across subspaces. Central to our framework is a trajectory-level objective that jointly trains the score network and learns $M_t(\theta)$, which encompasses general parameterization classes of matrix-valued noise schedules. We further derive an estimator for the derivative with respect to $\theta$ of the score that enables efficient optimization of the $M_t(\theta)$ schedule. For inference, we develop an efficiently-implementable reverse-ODE solver that is an anisotropic generalization of the second-order Heun discretization algorithm. Across CIFAR-10, AFHQv2, FFHQ, and ImageNet-64, our method consistently improves upon the baseline EDM model in all NFE regimes. Code is available at this https URL. Subjects: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV) Cite as: arXiv:2602.19512 [cs.LG] (or arXiv:2602.19512v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2602.19512 Focus to learn more arXiv-...
