跳到正文
arXiv cs.AI· Saptarshi Roy, Alessandro Rinaldo, Purnamrita Sarkar·· 2 小时前AI 评分41

从扩散与随机定位视角研究 Rectified Flow 的低维自适应采样加速

Low-Dimensional Adaptation of Rectified Flow: A Diffusion and Stochastic Localization Perspective

AI 导读

这篇 arXiv 论文分析 Rectified Flow(RF)如何自适应目标分布支撑的低维性以加速采样:在端点适配的时间离散化方案和足够精确的漂移估计下,RF 采样器的迭代复杂度为 O(k/ε)(至多差对数因子),ε 为总变差距离精度,k 为内在维度。

正文 · 原文

View PDF HTML (experimental)

Abstract:In recent years, Rectified flow (RF) has gained considerable popularity largely due to its generation efficiency and state-of-the-art performance. In this paper, we investigate how well RF automatically adapts to the intrinsic low dimensionality of the support of the target distribution to accelerate sampling. We show that, using a carefully designed end-point adapted time-discretization scheme and with sufficiently accurate drift estimates, the RF sampler enjoys an iteration complexity of order $O(k/\varepsilon)$ (up to log factors), where $\varepsilon$ is the precision in total variation distance and $k$ is the intrinsic dimension of the target distribution. In addition, we show that the denoising diffusion probabilistic model (DDPM) procedure is equivalent to a stochastic version of RF by establishing a novel connection between these processes and stochastic localization. Building on this connection, we further design a stochastic RF sampler that also adapts to the low-dimensionality of the target distribution under mild requirements on the accuracy of the drift estimates, and also with a specific time schedule. We illustrate the efficacy of newly designed time-discretization schedules with simulations on the synthetic data and text-to-image (T2I) data experiments.
Comments: 48 pages, 7 figures
Subjects: Machine Learning (stat.ML); Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Statistics Theory (math.ST)
Cite as: arXiv:2601.15500 [stat.ML]
  (or arXiv:2601.15500v5 [stat.ML] for this version)
  https://doi.org/10.48550/arXiv.2601.15500

arXiv-issued DOI via DataCite

Submission history

From: Saptarshi Roy [view email]
[v1] Wed, 21 Jan 2026 22:09:27 UTC (7,148 KB)
[v2] Tue, 3 Feb 2026 19:21:19 UTC (7,424 KB)
[v3] Fri, 20 Feb 2026 22:57:16 UTC (7,424 KB)
[v4] Sun, 1 Nov 2026 12:54:11 UTC (11,416 KB)
[v5] Wed, 18 Nov 2026 18:32:58 UTC (11,408 KB)

来源:arXiv cs.AI · arxiv.org