acceptodds
Withdrawn from ICLR 2027

EbbFlow: PINN-Inspired Progressive Curriculum for MeanFlow Training

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

MeanFlow learns interval-average velocity for few-step generation, but its fixed temporal sampler mixes diagonal and interior supervision throughout training. We interpret temporal coordinates as collocation points: diagonal samples provide a stochastic boundary target, whereas finite intervals supply a differential-consistency target. The diagonal fraction controls how often the consistency component appears, while interval length enters its gradient through an explicit factor. Motivated by this distinction, EbbFlow uses Temporal Supervision Allocation (TSA) to schedule the diagonal fraction and interval geometry. It starts with diagonal supervision, then admits interior samples and expands their maximum interval; a left-anchored interval law preserves coverage across time. The per-sample loss, target, backbone, and inference procedure remain unchanged. In matched class-conditional ImageNet runs without classifier-free guidance (DiT-B/2, one training seed per method, 400k updates), EbbFlow achieves 50k-sample FID of 39.65 at one network evaluation and 36.02 at two-evaluation ODE sampling, versus 44.82 and 39.86 for MeanFlow. Gradient probes show a mid-training difference in component geometry. Ablations support allocation scheduling and left-anchored intervals, although a fixed full-horizon variant performs better at one evaluation. These results suggest that temporal sampling is a useful training control whose benefits depend on the inference protocol.

Primary area:
generative models
Keywords:
MeanFlow, flow matching, temporal supervision allocation, curriculum learning, physics-informed neural networks, few-step generation

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_w9CeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the temporal coordinate sampling strategy in MeanFlow training. MeanFlow models the average velocity over an interval (r,t)(r,t) to enable few-step generation. The authors view the temporal interval (r,t)(r,t) as collocation points on a triangular domain and decompose the sampling process into a r=tr=t boundary-condition target and a r<tr<t interior-interval target. Based on this idea, the paper proposes Temporal Supervision Allocation (TSA), which first uses diagonal supervision to stabilize the boundary and then progressively introduces interior samples. While introducing interior samples, it uses a left-anchored interval law to control interval length and gradually expands from short intervals to the full interval. The method only changes the temporal sampling strategy and does not alter the rest of MeanFlow's training or inference pipeline. Experiments show certain advantages for few-step generation on DiT-B/2 with ImageNet.

Critical strengths

The problem definition is clear. It separates the two constants in MeanFlow's temporal sampling into two independent schedules, i.e., diagonal allocation and interval geometry, yielding a concise and easy-to-understand perspective. The method only modifies the sampling strategy without changing other components, making the algorithm simple to implement with low additional cost.

Critical weaknesses
  1. TSA actually optimizes a different weighted objective. Compared with fixed sampling, its gains may come from effective data sampling or schedule tuning rather than purely from a "progressive curriculum". From Table 2, replacing the interval sampling law from Min-Max with Left-Anchored (c→e) yields a solid performance improvement, whereas introducing the quadratic horizon schedule (e→f) instead leads to a degradation in 1-NFE FID.
  2. The experiments are relatively thin, and there are many hyper-parameters requiring manual tuning. Baseline comparisons are insufficient, AlphaFlow uses a different σ\sigma and target annealing. The evaluation setting is limited, all experiments are run only on the medium-scale DiT-B/2 (131M parameters) for 400k steps without classifier-free guidance (CFG), resulting in relatively high overall FID (between 36 and 44), which leaves a substantial gap from practical generation quality.
Questions for authors
  1. To validate the FID improvements of EbbFlow, it is recommended to use 3–5 training seeds and report the mean results with standard deviation of EbbFlow and baselines.
  2. Regarding hyper-parameter robustness. How sensitive is performance to ρend\rho_{end}​, γ\gamma, and ⁡hmax⁡h_{max}​?
Reviewer_GukAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes EbbFlow, a temporal sampling curriculum for MeanFlow training. Inspired by the distinction between boundary and interior constraints in physics-informed neural networks, it schedules the fraction of diagonal samples, introduces a left-anchored interval distribution, and progressively expands the maximum interval length. The architecture, per-sample loss, target construction, and inference procedure remain unchanged.

Experiments use class-conditional ImageNet 256×256 with DiT-B/2, 400k training updates, and no classifier-free guidance. EbbFlow improves FID over the matched MeanFlow baseline from 44.82 to 39.65 at one network evaluation and from 39.86 to 36.02 at two-evaluation ODE sampling. Ablations and gradient probes investigate the contributions of sampling choices and changes in optimization geometry. Each training configuration uses one seed.

Critical strengths
  1. A simple intervention with useful improvements under matched conditions. EbbFlow modifies temporal sampling without increasing model size or inference cost. The improvements over matched MeanFlow and AlphaFlow baselines suggest that temporal supervision allocation is a useful training control. Its implementation simplicity makes the approach potentially practical.

  2. Transparent reporting and informative component comparisons. The paper explicitly acknowledges its single-seed evaluation, inference-protocol dependence, and the stronger one-step performance of the full-horizon control. The ablations provide evidence that allocation scheduling and the interval distribution affect generation quality. This transparency helps distinguish supported findings from the proposed interpretation.

Critical weaknesses
  1. The evidence does not establish the value of the complete progressive curriculum over simpler sampling alternatives. In Tables 2 and 5, removing horizon expansion while retaining scheduled allocation and left-anchored sampling improves one-NFE FID from 39.65 to 38.85 and two-NFE consistency FID from 39.97 to 39.37. The complete method improves two-NFE ODE FID by only 0.23, from 36.25 to 36.02, without training-seed uncertainty estimates. Moreover, the experiments omit a fixed-allocation, full-horizon, left-anchored baseline. This control is important because left anchoring substantially increases exposure near the one-step inference coordinate. Consequently, the results support modifying temporal sampling more strongly than they support the complete proposed curriculum. The gradient probes do not resolve this attribution issue: they are post-hoc diagnostics, do not follow the actual training sampling distribution, and do not include the strongest full-horizon control.

  2. The incremental contribution relative to existing curricula and its broader empirical significance remain insufficiently established. As acknowledged in Section 2, TEMF already schedules boundary-to-interior allocation, while Kim et al. study short-to-long interval learning. EbbFlow differs in its sampling construction and preservation of the baseline per-sample loss, but it does not provide matched comparisons with these closely related approaches. Validation is also restricted to one dataset, one backbone, one training budget, and an unguided setting, with one training seed per configuration. The reported absolute FID is substantially higher than published results in stronger configurations, although those numbers are not directly comparable. The concern is therefore not the raw cross-paper gap itself, but the lack of evidence that the proposed recipe remains useful in a more competitive setting. Given the overlap with prior curricula, stronger comparative evidence is needed to establish a sufficiently consequential contribution.

Questions for authors
  1. Evidence for horizon expansion: Do you have repeated-training results for the full-horizon left-anchored control and the complete method? Is the 0.23 FID improvement under two-NFE ODE sampling reproducible, and how should it be weighed against the deterioration under the other two inference protocols?

  2. Attribution to curriculum versus sampling distribution: Can you provide a fixed-allocation, full-horizon, left-anchored control? The existing fixed-allocation left-anchored variant still uses horizon expansion, so it does not establish whether a static sampler can achieve comparable performance.

  3. Distinction from the closest prior approaches: What evidence demonstrates an advantage over TEMF’s allocation scheduling or the short-to-long curriculum of Kim et al. under comparable conditions? Please distinguish differences in implementation from demonstrated benefits.

  4. Choice of configuration and robustness: How were the allocation endpoints, warm-start duration, and quadratic horizon selected, and what tuning budget was used for each method? Are there existing results showing that the gains persist with classifier-free guidance, a different model scale, or a longer training budget?

Reviewer_4zZWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EbbFlow, a PINN-inspired progressive curriculum for improving MeanFlow training. In this paper, the authors reinterpret MeanFlow's diagonal and interior supervision as boundary and differential-consistency constraints and introduce Temporal Supervision Allocation (TSA), which progressively adjusts the diagonal sampling ratio and interval length using a left-anchored sampling strategy. The method preserves the original training objective, architecture, and inference procedure. Experiments on ImageNet 256×256 demonstrate improved FID over MeanFlow and AlphaFlow in both one- and two-step generation, supported by ablation studies and gradient analyses.

Critical strengths

1. Principled Interpretation of MeanFlow Training through Temporal Collocation

The paper provides an interesting PINN-inspired perspective by interpreting MeanFlow's diagonal and interior supervision as boundary and differential-consistency constraints, respectively. This interpretation helps explain the different optimization characteristics of these supervision signals and provides a principled motivation for curriculum-based training. The gradient analyses further support this perspective by examining how gradient conflicts vary across interval lengths and training stages.

2. Flexible and Practical Temporal Supervision Scheduling

The proposed Temporal Supervision Allocation (TSA) framework jointly schedules two complementary factors: the diagonal sampling ratio and the interval horizon. Unlike prior approaches that primarily modify training targets or loss weighting, EbbFlow controls the temporal sampling distribution while preserving the original MeanFlow objective, architecture, and inference procedure. The left-anchored sampling strategy also maintains global temporal coverage while controlling interval lengths. This offers a practical and potentially complementary approach to improving MeanFlow training.

Critical weaknesses

1. Limited Experimental Validation and Relatively Weak Generation Quality

The experimental evaluation is limited to a single DiT-B/2 architecture on ImageNet 256×256, with 400K training updates and no classifier-free guidance. Although EbbFlow improves upon the matched MeanFlow and AlphaFlow baselines, the absolute generation quality remains relatively weak, with FID scores of 39.65 (1-NFE) and 36.02 (2-NFE).

In contrast, the original MeanFlow has demonstrated substantially stronger generation quality under more extensively trained and guidance-enabled settings (e.g., 1-NFE FID of 6.17 with MeanFlow-B/2). While these results are not directly comparable due to differences in training budgets and guidance settings, the large gap raises concerns about whether the proposed improvements would persist in more competitive, high-quality generation regimes.

Evaluations with larger backbones, longer training budgets, and standard guidance settings would be important to establish the generalizability and practical significance of the proposed method.

2. Limited Evidence for the Benefit of Joint Progressive Scheduling

Although jointly scheduling the diagonal allocation and interval horizon is presented as a central contribution, the ablation results do not consistently demonstrate the advantage of the complete curriculum. In Table 2, the fixed full-horizon left-anchored variant achieves better 1-NFE FID (38.85 vs. 39.65), while the complete EbbFlow improves 2-NFE FID by only 0.23 (36.25 vs. 36.02).

Moreover, all experiments are conducted with a single training seed, making it difficult to assess whether such a small improvement is statistically reliable. These results suggest that much of the performance gain may come from diagonal allocation scheduling and the left-anchored sampling distribution, rather than the proposed progressive interval horizon. Stronger evidence is needed to establish the additional benefit of jointly scheduling both factors.

Questions for authors

N/A

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

Related papers

Discussion (0)

Sign in to comment.