---
title: "优化你的采样：用贝叶斯优化调优扩散模型采样调度"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.18040v1"
permalink: "https://feed.astralor.com/items/itm_3c7f0c7b652383c510a06b2c"
published: "2026-08-18T17:33:00.000Z"
score: 77
channel: "ai"
tags: ["OYS","扩散模型","文生图","贝叶斯优化","采样调度"]
---

# 优化你的采样：用贝叶斯优化调优扩散模型采样调度

## AI 点评

> 这篇论文值得关注，因为它跳出了以往“改进求解器”的套路，直接对采样时间步做黑盒优化，思路简洁却效果惊人。对正在做扩散模型推理加速的工程师和研究者来说，OYS几乎可以零成本接入现有流程，尤其适合追求低延迟的落地场景。

## 速览

从扩散模型采样通常需要多次前向传播，计算成本高昂。现有工作大多关注高效求解器或采样器，却很少关注采样时间步本身的选择。近期研究多通过理论推导替代指标来优化质量，而非直接优化真实指标。本文提出OYS方法，将时间步选择视为黑盒优化问题，直接用贝叶斯优化优化目标指标。OYS在文生图任务上优于默认调度和Align Your Steps，在内绘及其他图像任务上也优于默认调度，定量与人工评估均验证其优势。该方法无需额外训练，适用于蒸馏模型，并能增强Euler和DPM-Solver++等高级采样器。5步OYS调度可保留50步调度的89%-94%质量，同时推理成本降低10倍。

## 关键点

- OYS将扩散模型时间步选择视为黑盒优化，直接以目标指标为优化目标，无需额外训练。
- 在文生图、内绘等任务上，OYS生成的采样调度优于默认调度及Align Your Steps方法。
- 5步调度可保留50步调度89%-94%的图像质量，推理成本降低10倍。
- OYS适用于蒸馏模型，并能同时改进Euler与DPM-Solver++等采样器。

## 全文

Sampling from a diffusion model typically requires many forward passes through a large neural network, making generation computationally expensive. While much work has focused on efficient solvers and samplers, comparatively little attention has been paid to selecting the sampling timesteps themselves. A recent line of work optimizes theoretically derived surrogates for sample quality rather than the quality metric itself. We propose Optimizing Your Sampling (OYS), which instead treats timestep selection as a black-box optimization problem, optimizing the target metric directly with Bayesian optimization. OYS outperforms both the default schedules and those of Align Your Steps on text-to-image generation, and improves over the default schedules on inpainting and other image tasks, in both quantitative and human evaluations. OYS requires no additional training, is applicable even to distilled models, and improves both simple and sophisticated samplers such as Euler and DPM-Solver++. A 5-step OYS schedule retains 89%-94% of the quality of a 50-step schedule while reducing inference cost by 10x.

---

引用前请核对原文;本文由星际流动整理:https://feed.astralor.com/items/itm_3c7f0c7b652383c510a06b2c