acceptodds
Under review as a conference paper at ICLR 2027

Rectified MeanFlow: Overcoming the Curvature Bottleneck in MeanFlow

28 views

Abstract

Flow-based models are powerful generators, but sampling typically requires costly numerical integration along curved generative trajectories. MeanFlow offers a promising framework to few-step generation by directly learning mean-velocity fields. However, we find that the highly curved generative trajectories of existing models induce a noisy loss landscape, severely bottlenecking convergence and model quality. We leverage a fundamental geometric principle to overcome this: mean-velocity estimation is drastically simpler along straight paths. Building on this insight, we propose Rectified MeanFlow, a self-distillation approach that learns the mean-velocity field over a straightened velocity field, induced by rectified couplings from a pretrained model. By smoothing the optimization landscape, our method achieves strong few-step generation performance. Under the same pretrained initialization and matched end-to-end post-training budgets, including coupling generation, Re-MeanFlow reduces guided one-step FID from to on ImageNet-. In a controlled text-to-image proof of concept using Cosmos-Predict2-2B, it improves GenEval from to at NFE, compared with approximately for the -NFE teacher. Together, these results highlight the synergy between trajectory rectification and mean-velocity modelling as an effective route to compute-efficient few-step generation.

Primary area:
generative models
Keywords:
Flow Matching, Efficient Training, Efficient Generation, MeanFlow

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_X9YROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents Rectified Meanflow, a framework to improve one- and few-step generation by training Meanflow on rectified noise image pairs generated by a pretrained generative model. The paper claims that straighter transport trajectories simplify mean velocity learning, leading to faster convergence and better sample quality. Experiments on ImageNet and text-to-image generation show substantial improvements over direct Meanflow fine-tuning.

Critical strengths
  • Clear presentation of the benefit of coupling

The authors show that MeanFlow trained on the same generated images with independent noise barely beats Meanflow on real data. That shows the gain comes from keeping the noise–sample correspondence, not from the synthetic images themselves.

  • Interesting presentation

Showing the loss as a function of the input z_t for different (r, t) pairs is an intuitive way to illustrate the problem. As the jump t − r grows, the Meanflow landscape gets visibly sharper while the Re-MeanFlow landscape stays smooth. It makes the motivation of the paper easy to follow and demonstrates where the gains might come from.

Critical weaknesses
  • Curvature claim not isolated

The authors claim that reduced curvature is the actual reason for the gains, but the rectified coupling also makes the regression target almost noise-free, which can explain the faster convergence just as well. AYF already avoids noisy targets by using the teacher velocity, but the comparison isnt fair, since AYF doesnt use the proposed training tricks, which alone give a large improvement as table 4 shows. If the authors provide AYF trained with the same insights and a loss-landscape plot that shows AYF landscape is not smooth, the argument would be much more convincing. Especially since AYF is very close in terms of FID.

  • Gain in other baselines

Does the gain also hold for other baselines such as Improved Meanflow? I am not asking about the largest model, rather, I am wondering whether the rectified coupling is a general plug-in that also helps the newer MeanFlow variants, or whether the improvement mostly disappears once the objective and guidance are already improved, as in iMF. Even a small-scale experiment, e.g. iMF-B at 256x256 with and without rectified couplings under the same budget, would answer this.

Questions for authors
  • Is the loss-landscape of AYF smooth as well?
  • Does the method still beat AYF when training design of table 4 is applied to AYF?
  • Is the gain transferable to SOTA?
Reviewer_Ff2KOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues that MeanFlow is hard to optimize because the independent noise–data coupling induces highly curved trajectories, which makes the mean-velocity regression target harder to learn. It proposes Re-MeanFlow. The method generates one round of reflow pairs with a pretrained flow or diffusion model, initializes the student from the same model, and trains the standard MeanFlow objective on these pairs. It adds a few training choices: time conditioning that preserves the initialization, milder loss normalization, modified time sampling, and redirecting a high-noise, short-interval region to the clean endpoint. The method uses no real data. The curvature argument is supported by a correction-gap analysis, a curvature proxy and input-space loss landscapes. On ImageNet, the method achieves better FID and faster convergence than the baselines with little compute, and it is further validated on text-to-image generation.

Critical strengths
  1. The ImageNet ablation shows that the improvement comes from preserving the noise–sample pairing, rather than simply using synthetic images. The MeanFlow baseline uses the same training recipe, and the compute comparison includes the cost of generating the pairs. Consistent gains across resolutions and backbones support the benefit of the proposed pairing.

  2. The method requires no real data, auxiliary networks, or adversarial or score-distillation losses. It achieves one-step quality comparable to recent distillation methods, with lower estimated compute.

Critical weaknesses
  1. Generating noise–image pairs with a pretrained diffusion model follows the same procedure used for ODE-pair regression in DMD (Yin et al., One-Step Diffusion with Distribution Matching Distillation, CVPR 2024). Although the paper cites DMD, it does not explicitly discuss this connection, which makes the novelty of the data-generation procedure appear overstated. I still see value in applying these pairs to MeanFlow, which, to my knowledge, has not been explored before. The authors should distinguish this contribution from the existing ODE-pair construction.

  2. My main concern is the distillation baseline. Since the method requires a pretrained diffusion model to generate training pairs, it should be evaluated as a distillation method. Table 2 initializes the MeanFlow baseline with teacher weights, but initialization alone does not fully use the teacher. In MeanFlow distillation, a frozen teacher can provide the instantaneous velocity, while the student learns the average velocity. I would like to see this baseline, with both teacher initialization and teacher-velocity supervision, and an ablation comparing ODE pairs against independently paired real data. This would help establish whether the proposed pairing remains beneficial under a stronger distillation baseline.

  3. Faster convergence comes with the cost of generating training pairs through multi-step teacher sampling. Table 2 includes this cost, which is appropriate, but data generation still takes roughly half of the reported total budget. The paper should distinguish the reduction in student training time from the overall computational savings.

  4. Although the main claim concerns convergence speed, longer training would help clarify whether the method also improves final quality. For example, under the 240-epoch setting used by Improved MeanFlow, does Re-MeanFlow still outperform the baseline, or does the baseline eventually close the gap?

Questions for authors
  1. Could you compare against MeanFlow distillation using teacher weight initialization and frozen teacher velocities, and ablate ODE pairs versus independently paired real data?

  2. Does the advantage persist with longer training, or does the baseline eventually close the gap? Please report total compute, including pair generation.

Reviewer_CPavOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper combines Rectified Flow with MeanFlow and proposes Rectified MeanFlow, which uses reflow to address the problem of a noisy loss landscape induced by highly curved generative trajectories in MeanFlow. The method is validated on ImageNet at multiple resolutions, with a simple text-to-image proof of concept.

Critical strengths
  1. The proposed method is data-free.
  2. Under matched end-to-end post-training budgets, the proposed method achieves substantial performance improvements.
  3. The ablation study thoroughly validates the effects of each training component and of the coupling geometry.
Critical weaknesses
  1. The paper is closer to combining two existing works, so its novelty is relatively limited.
  2. It is recommended to validate on different datasets. The current main experiments cover only three ImageNet resolutions.
  3. The performance gaps in Table 1 are small. Multiple runs with standard deviations or confidence intervals are recommended.
Questions for authors
  1. Please see weaknesses.
  2. If multiple reflow steps are performed, how do the performance change?
  3. The authors acknowledge in Appendix A.1 that real data can be incorporated. It is recommended to add results on real data where available, or compare under the same data condition.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.