Flow Straighter and Faster: Efficient One-Step Generative Modeling via MeanFlow on Rectified Trajectories (2511.23342v1)

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.

• The paper introduces Re-Meanflow, a hybrid model combining Rectified Flow and MeanFlow to achieve one-step generative modeling with significantly reduced computational cost.
• It leverages trajectory rectification and mean velocity modeling to stabilize training and improve sample quality, as evidenced by a 33.4% FID score improvement on ImageNet.
• Its efficient design democratizes access to high-quality generative models, encouraging applications in diverse fields like image processing and bioinformatics.

Efficient One-Step Generative Modeling via Re-Meanflow

Rectified MeanFlow (Re-Meanflow) presents a novel approach for one-step generative modeling by combining the strengths of Rectified Flow and MeanFlow methodologies to enhance both sampling efficiency and generative quality. The problem of high computational cost associated with generating high-quality samples in flow-based and diffusion generative models lays the foundation for this paper, which proposes a more efficient hybrid model achieving significant performance improvements.

Background and Motivation

Flow-based models offer superior generative modeling capabilities, though they traditionally require computationally expensive iterative solutions like ordinary differential equation (ODE) integration. Rectified Flow attempts to reduce these challenges by straightening trajectories across iterative reflow steps. However, achieving near-perfect straightness demands computational resources that are economically impractical.

MeanFlow suggests a paradigm shift by modeling the mean velocity of trajectories, avoiding the need for perfectly straight paths. However, its training becomes unstable and noisy on highly curved trajectories, reducing convergence speed.

Re-Meanflow strategically combines both approaches, leveraging trajectory rectification to simplify MeanFlow training, thereby eliminating the requirement for multiple reflow steps while also achieving higher fidelity in generated samples.

Methodology

Re-Meanflow operates by modeling the mean velocity over rectified trajectories, facilitated by a single reflow step prior. This hybrid method harnesses the straightened trajectory outputs from a single-step reflow to form a more stable basis for MeanFlow training. More stable and less variably curved paths during training result in less noise and higher sample quality, enhancing convergence efficiency.

The framework introduces a truncation heuristic to manage curvature extremes by discarding a specified percentage of the highest curvature couplings. This decision aims to minimize variability caused by significant curvature, thus leading to more consistent generative quality across samples.

Figure 1: Why trajectory rectification and mean-velocity modeling reinforce each other. (a) A 1-rectified flow still follows highly curved trajectories, (b) MeanFlow estimates become complex and challenging to learn in high curvature fields.

Experimental Results

Extensive evaluations on ImageNet datasets at multiple resolutions ($64^2$, $256^2$, and $512^2$) indicate that Re-Meanflow outperforms prior distillation methods and baseline counterparts such as 2-rectified flow++, improving Fréchet Inception Distance (FID) scores by 33.4% while utilizing merely 10% of related computational resources.

Figure 2: Comparison of total computational cost on ImageNet-$64^2$ showcasing reduced compute requirements with Re-Meanflow.

Re-Meanflow demonstrated enhanced quality versus computational costs and exhibited stable training stages, resulting in high-quality one-step sampling. Its architecture design maximizes the quality-to-compute ratio by shifting burdens to widely accessible inference-stage computations.

Implications and Future Directions

Re-Meanflow has established efficacies demonstrating potential beyond current generative model frameworks. By circumventing the prohibitive computational needs for training few-step or one-step distillations, Re-Meanflow provides a more accessible entry point into high-quality generative modeling, democratizing its applications in machine learning, bioinformatics, and image processing.

Future work could explore incorporating real data into training to leverage synthetic data synergies, or expanding Re-Meanflow's principles to broader generative tasks such as text-to-image or multimodal generation contexts.

Conclusion

Re-Meanflow introduces a robust, computation-efficient generative modeling technique that significantly enhances training efficiency and sample quality. By capitalizing on the synergy between rectified flows and mean velocity modeling, it positions itself as a promising solution for real-world application scalability in generative model training.

Figure 3: Qualitative results for Re-Meanflow (NFE=1) on ImageNet 64^2, 256^2, and 512^2.

The methodology and results collectively suggest that adopting mean-flow modeling over simplified trajectory paths can potentially open new avenues in efficient generative model training and applications.