Flow Straighter and Faster: Efficient One-Step Generative Modeling via MeanFlow on Rectified Trajectories

Agentic AI
Published: arXiv: 2511.23342v1
Authors

Xinxi Zhang Shiwei Tan Quang Nguyen Quan Dao Ligong Han Xiaoxiao He Tunyu Zhang Alen Mrdovic Dimitris Metaxas

Abstract

Flow-based generative models have recently demonstrated strong performance, yet sampling typically relies on expensive numerical integration of ordinary differential equations (ODEs). Rectified Flow enables one-step sampling by learning nearly straight probability paths, but achieving such straightness requires multiple computationally intensive reflow iterations. MeanFlow achieves one-step generation by directly modeling the average velocity over time; however, when trained on highly curved flows, it suffers from slow convergence and noisy supervision. To address these limitations, we propose Rectified MeanFlow, a framework that models the mean velocity field along the rectified trajectory using only a single reflow step. This eliminates the need for perfectly straightened trajectories while enabling efficient training. Furthermore, we introduce a simple yet effective truncation heuristic that aims to reduce residual curvature and further improve performance. Extensive experiments on ImageNet at 64, 256, and 512 resolutions show that Re-MeanFlow consistently outperforms prior one-step flow distillation and Rectified Flow methods in both sample quality and training efficiency. Code is available at https://github.com/Xinxi-Zhang/Re-MeanFlow.

Paper Summary

Problem
Generative models are powerful tools for creating synthetic data, but they often rely on expensive numerical integration of ordinary differential equations (ODEs) to sample new data points. This can be a major bottleneck in training and using these models. Researchers have been exploring ways to make this process more efficient, but existing solutions often require multiple iterations or sacrifice some level of performance.
Key Innovation
The authors of this paper propose a new framework called Rectified MeanFlow (Re-MeanFlow) that addresses these limitations. Re-MeanFlow combines the strengths of two existing approaches: Rectified Flow and MeanFlow. It learns the mean velocity field along rectified trajectories, which are obtained after a single reflow step. This allows for efficient training and improves the quality of generated data.
Practical Impact
The practical impact of Re-MeanFlow is significant. By enabling efficient training on large datasets, it can help researchers and practitioners create more accurate and realistic synthetic data. This can have applications in a wide range of fields, such as computer vision, natural language processing, and data augmentation. Additionally, Re-MeanFlow can be used to scale one-step generative paradigms to larger domains, such as text-to-image generation.
Analogy / Intuitive Explanation
Imagine you're trying to navigate a winding road to get to your destination. The traditional approach would be to try to drive the entire road at once, which can be difficult and time-consuming. Re-MeanFlow is like taking a shortcut by straightening out the road, making it easier to drive and reducing the risk of getting lost. By learning the mean velocity field along these straightened trajectories, Re-MeanFlow can efficiently generate new data points that are more accurate and realistic.
Paper Information
Categories:
cs.CV cs.AI
Published Date:

arXiv ID:

2511.23342v1

Quick Actions