Flow Straighter and Faster: Efficient One-Step Generative Modeling via MeanFlow on Rectified Trajectories
Xinxi Zhang, Shiwei Tan, Quang Nguyen, Quan Dao, Ligong Han, Xiaoxiao He, Tunyu Zhang, Alen Mrdovic, Dimitris Metaxas
2025-12-02
Summary
This paper introduces a new method called Rectified MeanFlow for generating images. It builds on previous techniques that use 'flows' to create images, aiming to make the process faster and produce higher quality results.
What's the problem?
Existing methods for generating images using flows often require a lot of computational power. Some methods need to carefully straighten the paths the data takes during generation, which takes many steps. Others try to directly predict the average path, but struggle when the underlying flow is complex, leading to slow learning and blurry images.
What's the solution?
Rectified MeanFlow solves this by predicting the average path along a slightly straightened trajectory, but only needing a single straightening step. This is more efficient than methods requiring many steps to fully straighten the path. They also added a simple trick to further reduce any remaining curves in the path, improving the final image quality.
Why it matters?
This research is important because it makes image generation using flows more practical. By reducing the computational cost and improving the quality of generated images, it opens the door for using these techniques in more applications and with higher resolution images, as demonstrated by their success on the ImageNet dataset at various resolutions.
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.