acceptodds
Under review as a conference paper at ICLR 2027

MixFlow2: One-Stage MixFlow Training for Flow Matching

Abstract

MixFlow reduces the training–testing discrepancy by training the velocity prediction network at a training timestep using training noisy data formed by teacher noisy data at and higher-noise teacher noisy data at slowed timesteps. It is different from standard training: training noisy data is only from the teacher noisy data at the training timestep . The original MixFlow approach adopts a two-stage training strategy: standard training first learns accurate teacher-forced velocity prediction, and MixFlow post-training then uses higher-noise teacher forcing to handle the Slow Flow problem in multi-step discrete sampling. Direct application of original implementation to one-stage training, i.e., training from scratch, is not optimal. The reason is that training noisy data sampling is more from higher-noise teachers, and more capable of handling the Slow Flow problem and less capable of learning teacher-forced velocity prediction. We present a one-stage training approach, MixFlow 2, with the introduction of a novel sampling mechanism: the slowed timestep for forming the training noisy data is sampled from a symmetric triangular distribution, where intermediate slow timesteps are sampled more frequently. In addition, we observe that the prediction at high-noise training timesteps is less reliable, as high-noise timestep is less-frequently sampled. This influences guided sampling: guidance at high-noise timesteps is harmful. So we adopt guidance interval to remove the guidance at high-noise timesteps, maintaining the generation quality achieved without guidance. Empirical results show that MixFlow 2 improves the class-conditional and text-to-image generation performance and outperforms original MixFlow.

Primary area:
generative models
Keywords:
Flow Matching

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_TrcvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes MixFlow2, a one-stage variant of the original two-stage MixFlow procedure. The main change is to replace the original noise-skewed training distribution with a symmetric triangular distribution, aiming to balance instantaneous prediction and robustness to the proposed Slow Flow problem. The method also uses a restricted guidance interval during sampling.

Critical strengths

MixFlow2 is simple and may reduce training cost by avoiding the additional post-training stage required by the original MixFlow implementation.

Critical weaknesses
  1. The central design deliberately introduces a mismatch between the training input and its associated conditioning and target. Although the paper motivates this choice through the proposed Slow Flow problem, the justification is not sufficiently convincing. The resulting objective departs from the standard flow-matching training formulation, and the paper gives little theoretical or empirical evidence that this mismatch leads to a generally valid learning objective.

  2. The method concentrates training samples in selected intervals and then relies on an additional guidance strategy during sampling. This combination introduces several design choices and hyperparameters, and its robustness across models and sampling configurations is not well established. The current evidence does not show that the proposed distribution is generally preferable rather than a task-specific choice.

Questions for authors

Please respond to the weaknesses above. In addition, the paper cites almost no recent work, and the related-work discussion should be updated accordingly.

Reviewer_xh3yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MixFlow 2 enables one-stage MixFlow training by reducing the slowness parameter from 0.8 to 0.5. This changes the input-noise distribution to a symmetric triangular distribution. Experiments on ImageNet and text-to-image generation show improvements over standard FM and comparable quality to two-stage MixFlow with fewer training epochs.

Critical strengths
  • The one-stage recipe simplifies training. It achieves quality comparable to the original MixFlow with 800 rather than 1,000 epochs.

  • Improvements over standard FM are consistent across several model scales, with supporting text-to-image results.

Critical weaknesses
  • The contribution over MixFlow is modest. The objective is unchanged, and the main modification is tuning an existing parameter. The practical benefit is useful, but the conceptual advance appears limited.

  • The explanation for the improvement is not isolated experimentally. Changing the parameter affects both the input-noise distribution and the time mismatch. It remains unclear which change drives the gains.

Questions for authors
  • Can the input-noise distribution and time mismatch be varied independently to identify the source of improvement?

  • Is there direct evidence that the revised recipe better balances velocity learning and handling Slow Flow, beyond final generation scores?

Reviewer_iwvtOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MixFlow 2, a novel training timestep sampling schedule for flow matching models. The derivation of MixFlow 2 is motivated by observations of the original two-stage training of MixFlow. Experimental results on image generation tasks show the advancement of MixFlow 2 over MixFlow.

Critical strengths
  • The paper is clearly written and easy to follow.
  • The figure on different time-step sampling is illustrative for understanding.
  • The experimental results on image generation are abundant.
Critical weaknesses

(My primary concern) Some papers point out that we should look at the time-step sampling schedule in an information-theoretic way; we should adjust the schedule based on entropy, as shown in [1]. Their view seems more reasonable and has theory to back it up; the method in this paper is rather heuristic. In commercial generative model training, we would further adjust the schedule based on modality and token length. I think further comparison to a wider range of time schedules is needed.

[1] Ambrogioni, Luca, et al. "From Atoms to Entropy: Optimal Noise Allocation for Diffusion Training in the Convex Regime." arXiv preprint arXiv:2607

Questions for authors

see weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.