Go to ICLR 2026 Conference homepageRe-Meanflow: Efficient One-Step Generative Modeling via Meanflow on Rectified Trajectories
Keywords: Generative Model, Flow Matching, Efficient Training
TL;DR: RE-Meanflow combines the efficiency of Meanflow with the trajectory straightening of Rectified Flow to enable high quality one-step sampling with only 1 reflow process.
Abstract: Flow models have demonstrated remarkable capabilities in generative modeling, yet a key bottleneck is the expensive numerical integration of ODEs required for sampling. Rectified Flow mitigates this by iteratively learning straighter trajectories through a reflow mechanism, but repeated rounds of training incur heavy computation overhead and often degrade sample quality. More recently, Meanflow has shown strong one-step generation by directly modeling the average velocity across time, but training it from scratch is costly, as it must learn from noisy signals induced by highly curved flows. To address these limitations, we propose Re-Meanflow, which trains a Meanflow model on trajectories straightened once using a single reflow step. The key insight is that “reflow” alone is too costly to achieve nearly straight paths for one-step sampling, while Meanflow can tolerate less-straight trajectories but is prohibitively expensive to train from scratch. By combining them, Re-Meanflow leverages their complementary strengths: a single reflow step produces sufficiently straight trajectories, enabling efficient Meanflow training without the performance degradation of repeated reflow processes. We evaluate Re-Meanflow on ImageNet $64^2$ and $256^2$, where it achieves competitive or superior one-step generation compared to state-of-the-art methods while offering substantial training efficiency. In particular, on ImageNet $64^2$, our method improves the FID of 2-rectified flow++ by $33.4$% while reducing training cost by $90$%.
Primary Area: generative models
Submission Number: 8062
Loading