---
id: 20260904-T0-12
title: "CAT-Flow：曲率自适应步长，提速流匹配生成模型"
title_en: "CAT-Flow Speeds Up Flow Matching with Curvature-Adaptive Steps"
url: https://ai.daily.yangsir.net/daily/20260904-T0-12
issue_date: 2026-09-04
publish_date: 2026-09-03T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2609.01746
---

# CAT-Flow：曲率自适应步长，提速流匹配生成模型

arXiv新论文提出CAT-Flow（Curvature-Adaptive sTeps for Flow Matching），针对FLUX和Stable Diffusion 3.5等流匹配模型在ODE采样迭代过程中的效率瓶颈，通过曲率自适应步长策略减少必要采样步数。该方法在不明显牺牲生成质量的前提下加快推理速度。

## English Version

**CAT-Flow Speeds Up Flow Matching with Curvature-Adaptive Steps**

A new arXiv paper introduces CAT-Flow, a curvature-adaptive stepping method for flow matching models. It tackles the efficiency bottleneck in ODE-based iterative sampling used by systems like FLUX and Stable Diffusion 3.5. By adjusting step size according to trajectory curvature, CAT-Flow reduces required sampling steps with minimal quality loss.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2609.01746)

**详情页**：https://ai.daily.yangsir.net/daily/20260904-T0-12

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*