跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2602.19512 · v2

Variational Trajectory Optimization of Anisotropic Diffusion Schedules

首次提交 2026-02-23 12:56
版本更新 2026-11-15 10:51
雷达观测 2026-11-17 19:15
arXiv ↗PDF ↗

摘要

We introduce a variational framework for diffusion models with anisotropic noise schedules parameterized by a matrix-valued path Mt(θ)M_t(θ) that allocates noise across subspaces. Central to our framework is a trajectory-level objective that jointly trains the score network and \emph{learns} Mt(θ)M_t(θ), which encompasses general parameterization classes of matrix-valued noise schedules. We further derive an estimator for ∂θ∇log⁡pt\partial_θ\nabla \log p_t that enables efficient optimization of the Mt(θ)M_t(θ) 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 FID with largest gains at low NFE. Code is available at https://github.com/kilonova801/anisotropic-diffusion.

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

本站首次收录为 v2,暂无更早版本可对照。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v2 · 2026-11-15 10:51 · 最新观测版本

Variational Trajectory Optimization of Anisotropic Diffusion Schedules

We introduce a variational framework for diffusion models with anisotropic noise schedules parameterized by a matrix-valued path Mt(θ)M_t(θ) that allocates noise across subspaces. Central to our framework is a trajectory-level objective that jointly trains the score network and \emph{learns} Mt(θ)M_t(θ), which encompasses general parameterization classes of matrix-valued noise schedules. We further derive an estimator for ∂θ∇log⁡pt\partial_θ\nabla \log p_t that enables efficient optimization of the Mt(θ)M_t(θ) 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 FID with largest gains at low NFE. Code is available at https://github.com/kilonova801/anisotropic-diffusion.

查看此版本解读 →查看此版本原文 ↗