acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Aware Time Reparameterization for Flow-Map Distillation

Abstract

Flow-map distillation enables one- and few-step generation by learning finite-time transitions of a pretrained generative ODE. We investigate whether changing the teacher's time parameterization can make these transitions easier to learn. Motivated by the hypothesis that trajectory segments with large normal acceleration are harder to distill, we propose a geometry-aware time reparameterization that allocates more student time to these regions while preserving the teacher's geometric paths and terminal distribution. We derive a shared clock that equalizes a population normal-acceleration statistic under suitable assumptions, and construct a practical approximation from robust, regularized estimates across teacher trajectories. We incorporate this clock into Lagrangian flow-map distillation, using the transformed time coordinate to condition the student. The clock is estimated once before distillation and requires neither teacher retraining nor additional student parameters or inference-time network evaluations. Experiments on synthetic data, CIFAR-10, and CelebA-64 show improved sample quality over identity-time distillation at matched inference budgets, including improvements in one- and two-step image generation. The gains in one-step generation, where no intermediate sampling times can be adjusted, highlight the benefits of time reparameterization during distillation.

Primary area:
generative models
Keywords:
Flow Matching, Consitency Models, Flow Map Matching, Flow Map Distillation

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_iCrxOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper asks whether flow-map distillation gets easier by only changing the time parameterization of a frozen teacher ODE. It estimates a normal-acceleration profile from teacher trajectories, builds a shared clock whose density follows the square root of this profile, and plugs the reparameterized field into Lagrangian flow-map distillation. Under the stated assumptions the clock equalizes the normal-acceleration statistic. On a 2D checkerboard, CIFAR-10 and CelebA-64, one and twostep FID to target improve over the identity clock, while results at more steps mostly get worse.

Critical strengths
  1. The clock keeps the teacher's paths fixed, is built offline and adds no network evaluations at sampling time, so it could be dropped into existing pipelines.
  2. The internal comparison shows gains at one and two steps. With matched architecture, budget and hyperparameters, one-step FID at 400k goes from 8.57 to 7.44 on CIFAR-10 and from 6.19 to 5.58 on CelebA-64. This shows the time parameterization matters, but not why.
Critical weaknesses
  1. first of all, the geometric signal is not separated from simpler factors. In the image experiments the objective can be rewritten as standard LMD with resampled time pairs, a residual weight and a transformed time input. This does not rule out geometry, but there is no identity-clock student trained on the same time-pair distribution, no normal versus full acceleration comparison, and no direct link between the profile and distillation error. So I think the key evidence for the main design choice is missing.
  2. further, the coverage and statistics are too limited to show practical value. Only CIFAR-10 and CelebA-64 are used, with two internal variants, no external few-step baselines, no image samples and no larger or latent-space settings. Everything is a single run, which matters most for the small two-step gains of 5.71 → 5.31 on CIFAR and 4.35 → 4.25 on CelebA at 400k. Results at 4, 8 and 20 steps mostly get worse, with one exception at 20 steps on CIFAR. I can't place the method's performance or the benefit of the geometric signal, and this is the main reason for my rejection.
Questions for authors

Please refer to major weakness. If the identity-clock student keeps physical time as input but samples time pairs from your clock's distribution, how much of the one-step gain remains? Under the same budget, what does the normal-acceleration clock gain over a full-acceleration clock, and is there any measurement linking per-time distillation error to the hardness profile? Could you report the number, means and variances of the preliminary repetitions, together with image samples? Why do results at more steps mostly degrade, and why do the CelebA two-step FID-to-target and the later teacher-relative metrics disagree? Do you have results for an identity-clock student using your sampling grid?

3 comments

We thank the reviewer for their detailed feedback. We have substantially revised the manuscript, particularly to address the concerns about experimental controls and the evidence supporting our geometric design choice.

1. Separating reparameterization from timestep sampling. We now include the requested identity-clock baseline trained with the same curvature-biased physical-time distribution, alongside a full 2x22 x 2 ablation (Section 5.2.3, Table 1). Importantly, curvature-biased sampling alone worsens performance, whereas reparameterization improves results under both sampling distributions. Our revised main experiments also use identical physical-time training distributions and inference grids, eliminating these sampling differences as explanations for the observed gains.

2. Stronger and more consistent results. With the revised protocol, the reparameterized student improves target FID at every evaluated inference budget (1, 2, 4, 8, and 20 NFEs) on both datasets, resolving the previously reported multi-step degradation relative to the identity baseline. Full FID/KID results across training checkpoints are provided in Appendix C. We also explicitly distinguish target fidelity from teacher-relative fidelity: improvements in the former do not necessarily imply improvements in the latter.

3. Theoretical motivation for normal acceleration. We have strengthened Section 4.1 and added Appendix D. We establish a local connection between normal acceleration and directional variation in finite-time average velocities, prove an optimality property of the equalizing clock, and analyze why exact full-acceleration equalization is more difficult and may not admit a smooth solution. While these results support our design choice, they do not establish a direct guarantee on neural-network distillation error.

4. Additional evidence. We have added qualitative CIFAR-10 and CelebA-64 samples (Appendix E, Figures 8–10) and measured the negligible offline construction overhead. We agree that additional training seeds, external baselines, and larger-scale evaluations would further strengthen the empirical evidence. These remain limitations of the current study.

We hope the new controlled experiments and theoretical analysis address the reviewer's principal concerns about soundness and clarify the contribution of time reparameterization beyond timestep sampling.

Thank you for the additional experiments. The sampling ablation addresses my concern that the gains might come from sampling high-curvature regions more often. I also appreciate the theoretical motivation for normal acceleration, though I would still like to see how it compares with full acceleration under the same training budget. The image experiments remain limited to CIFAR-10 and CelebA-64, so it is also difficult to judge how well the gains generalize. Given the stronger evidence in the rebuttal, I am raising my score from 1 to 2.

Thank you for taking the time to consider our rebuttal and for updating your assessment.

Regarding generalization, we have now completed additional experiments on class-conditional CIFAR-10, using a different pretrained flow-matching teacher (C2OT). The reparameterized student consistently improves target FID and KID across all evaluated inference budgets and training checkpoints. These results are included in the latest manuscript revision (Appendix C.4).

We acknowledge that a direct empirical comparison against a full-acceleration-based clock remains an important direction for further investigation. Our new Appendix D provides theoretical motivation for normal acceleration, but we do not claim that it establishes empirical superiority over alternative geometric diagnostics.

Thank you again for your constructive feedback and consideration.

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

This paper proposes a geometry-aware temporal schedule to improve flow map distillation. The authors hypothesize that trajectories with large normal acceleration are challenging to distill. To mitigate this difficulty, the authors propose that a learned temporal schedule which slows transport in high curvature areas can stabilize acceleration and improve generation quality. The authors construct this temporal schedule offline, using a dataset of pre-generated trajectories from a teacher model. Empirical results on toy and image generation demonstrate that flow maps distilled using this learned temporal schedule are generally more effective than those trained with the identity temporal schedule.

Critical strengths
  1. The experimental comparisons are well controlled. The authors use the same pretrained teacher and match training configurations for student learning, isolating the effect of the proposed temporal schedule.

  2. Results for the toy example demonstrate improved distributional fidelity (measured in KL Divergence,MMD and Wasserstein Distance) over the identity schedule for all evaluated sampling budgets.

Critical weaknesses
  1. Although the authors provide the schedule construction parameters, they do not report the cost of generating of the 256256 teacher trajectories with the adaptive ODE solver. It is difficult to assess the practical cost of constructing the schedule relative to the distillation benefit.

  2. For the image generation experiments, the reparameterized schedule improves one and two step generation, but underperforms the identity-time distillation for the four and eight step generations. The paper does not clarify why their proposed schedule appears less effective for multi-step generation.

Questions for authors
  1. For the image experiments, what are the adaptive ODE solver settings and wall-clock time required to generate the 256 teacher trajectories and to construct the schedule ϕ\phi?

  2. Why does the proposed schedule underperform identity-time distillation at the evaluated four and eight step budgets for the image generation experiments?

1 comment

We thank the reviewer for their constructive feedback and positive assessment. We have revised the manuscript to directly address both concerns.

1. Computational overhead and solver settings. We now use a fixed-step Heun solver with 100 steps (200 teacher evaluations) to generate 256 trajectories. Averaged over 10 runs on four V100 GPUs, the complete clock construction takes 9.57 s on CIFAR-10 and 53.66 s on CelebA-64, of which less than 40 ms is spent on acceleration estimation, aggregation, and interpolation. Relative to the original training budgets, we estimate a one-time overhead of only 0.013% and 0.015%, respectively. These estimates account for measured forward/training cost ratios, with details provided in Section 5.2 and Appendix C.1.

2. Performance at larger sampling budgets. We have revised the training procedure to use uniform physical-time sampling and now evaluate both students on identical physical-time inference grids. With these changes, the reparameterized student consistently outperforms identity-time distillation across all evaluated sampling budgets, including 4 and 8 NFEs, on both CIFAR-10 and CelebA-64 (Figure 5, Tables 2–3). The new sampling ablation (Section 5.2.3) further separates the effects of time reparameterization and timestep sampling.

We hope these additional measurements and experiments resolve the reviewer's concerns about practical overhead and multi-step performance.

Reviewer_5oRcOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a timestep reparameterization technique for few-step flow map distillation. Before training, it samples 256 teacher trajectories and estimate the normal-acceleration profile, and then compute a time transformation that equalizes the acceleration. The transformation reshapes teacher's flow map, on which the student is trained on. Experiments include only one baseline distillation framework (Lagrangian flow-map), on 2D toy dataset, CIFAR-10, and CelebA-64.

Critical strengths

The reparameterization method appears simple, intuitive, and potentially applicable to a broad range of trajectory-based distillation methods.

Critical weaknesses

Major concerns:

  • The experiments are insufficient.

    • A critical design choice of this work is normal acceleration, which distinguishes it from prior work like Sharpen Your Flow that used total acceleration. However, there is no ablations supporting the choice of using normal acceleration.
    • The experiments are limited to one baseline distillation framework (Lagrangian flow-map), and conduced under small-scale unconditional generation settings. No qualitative results are provided. And it appears that the proposed method is worse than the baseline at 4 and 8 NFEs.
  • The novelty of this work appears limited, and I feel that I learned very little from this paper. Timestep reparameterization is theoretically equivalent to a different student schedule plus time-dependent loss weighting. Both of these aspects have been explored extensively in prior work. This paper also does not isolate these underlying factors.

Relatively minor:

  • While normal acceleration appears to be a simple and intuitive metric, this paper does not have sufficient analysis that establishes the connection between normal acceleration and flow map distillation. Also it uses a static schedule, which does not account for trajectory-wise variations.

  • The practical significance of this work appears limited, since most large scale applications adopt distribution-based distillation instead of trajectory-based distillation. The benefit of curvature-based time warping may not be transferable to distribution-based methods.

Questions for authors

Can the authors provide comparisons against more baselines, such as different profile (e.g., total acceleration), or other weighting and time sampling strategies?

1 comment

We thank the reviewer for their feedback. We have substantially strengthened the manuscript to address the concerns regarding novelty, the choice of normal acceleration, and experimental evidence.

Reparameterization versus sampling schedules. In the revised experiments, we disentangle the effect of time reparameterization from timestep sampling through a new (2\times2) ablation (Section 5.2.3). We find that curvature-biased sampling alone degrades performance, while reparameterization consistently improves results under matched physical-time training distributions and inference grids. The updated experiments also show stronger performance across all evaluated inference budgets. These findings clarify that the benefits extend beyond changes in timestep allocation.

Normal versus full acceleration (new Appendix D). We now provide a theoretical motivation connecting normal acceleration to the leading directional correction in finite-time average velocities (Proposition D.1). We prove that our clock minimizes both the average and worst-time population normal-acceleration profile (Proposition D.2). We further show that, unlike normal acceleration, exact full-acceleration equalization requires solving a nonlinear differential constraint and may not admit a smooth solution (Propositions D.3–D.5).

Relation to Sharpen Your Flow. Although it uses full acceleration to construct an inference-time sampling schedule, its construction does not guarantee equalization or reduction of full acceleration under reparameterization. Our method instead constructs a clock with a provable normal-acceleration equalization property and applies it toflow-map learning rather than adjusting inference steps.

Together, these additions clarify the distinction from existing sampling methods, strengthen the geometric motivation, and provide stronger experimental support for our approach.

Comments on the paper

5 comments

We thank all reviewers for their detailed feedback. We have substantially revised the manuscript and conducted additional experiments to address two central concerns: the degradation at larger sampling budgets and the distinction between time reparameterization and nonuniform timestep sampling.

  1. Improved results at matched training and inference budgets

In the original submission, the reparameterized student was trained with uniformly sampled student times, inducing a nonuniform distribution over physical teacher times. We have now modified the training procedure to sample uniform physical teacher-time pairs ((s,t)) for both students, transforming only the coordinates supplied to the reparameterized student. This modification significantly improves the results. After 400k training updates, we obtain the following target FID results (Id-TP → ϕ\phi-TP):

CIFAR-10:

  • 1 NFE: 8.57 → 6.99
  • 2 NFE: 5.71 → 4.99
  • 4 NFE: 4.98 → 4.49

CelebA-64:

  • 1 NFE: 6.19 → 4.64
  • 2 NFE: 4.35 → 3.39
  • 4 NFE: 3.63 → 3.02

All results use the same training budgets, architectures, and identical uniform physical-time inference grids. The reparameterized student now consistently outperforms the identity student at all evaluated sampling budgets on both datasets, addressing the degradation at four and eight steps observed in the original submission. Importantly, the improvements in one-step generation cannot be attributed to inference-time scheduling, as both students use the same endpoints.

  1. Disentangling reparameterization from timestep sampling

To investigate whether these improvements simply arise from a different training-time distribution, we additionally conduct a complete (2\times2) ablation, combining: Identity or reparameterized student coordinates. Uniform or curvature-biased physical-time training sampling. The results (new Table 1, Section 5.2.3) show that: Curvature-biased sampling alone degrades performance (for both time parameterizations): identity-student one-step FID increases from 8.57 to 10.47 on CIFAR-10 and from 6.19 to 7.59 on CelebA-64. Reparameterization improves performance under both sampling distributions. The best results are obtained by combining reparameterization with a uniform physical-time sampling, which are the results reported in figure 5.

These observations show that the improvements cannot be explained by preferentially sampling high-curvature regions. Furthermore, our experiments employ EDM2-style adaptive loss weighting, which can compensate for time-dependent residual scaling. Since the underlying teacher flow map remains unchanged, these findings suggest that the transformed time coordinate induces more favorable finite-capacity approximation or optimization dynamics, potentially through an implicit inductive bias in the student's time-conditioning architecture.

  1. Manuscript revisions

We have revised the introduction to clarify these distinctions, updated the main experimental results and figures, and added the complete sampling-versus-reparameterization ablation. We acknowledge that these experiments do not establish if normal acceleration is the optimal geometric diagnostic. We will additionally report the computational overhead of constructing the clock, qualitative results, normal vs full acceleration clarification, along with results from ongoing conditional-generation experiments, in subsequent responses.

Note: The appendix tables have not yet been updated and still report results obtained using different physical-time inference grids for the identity and reparameterized students. The main-text results use identical uniform physical-time inference grids for both student models, as described above. We will update the appendix accordingly in the next revision.

We thank the reviewers again for their constructive feedback. Following our first rebuttal update, we have further revised the manuscript to address three additional concerns:

  1. Theoretical motivation for normal acceleration (Section 4.1, Appendix D). We provide new results connecting normal acceleration to the directional variation of finite-time average velocities, establish an optimality property of normal-acceleration equalization, and analyze the differences between normal and full acceleration under time reparameterization.

  2. Computational overhead (Section 5.2, Appendix C.1). We benchmark the complete offline clock-construction procedure. Its estimated one-time cost is only 0.013% of student training on CIFAR-10 and 0.015% on CelebA-64.

  3. Qualitative generation results (Appendix E, Figures 8–10). We provide generated samples for both datasets at 1, 2, and 4 NFEs, comparing the teacher, identity-time student, and reparameterized student using matched initial noise seeds.

The updated manuscript also includes the complete FID/KID tables using identical physical-time inference grids for both students, as described in our previous response.

We provide further details relevant to each review below.

Area_Chair_K2dROpenReview ↗

Please read the response from authors and discuss

Dear reviewer fellows,

The author rebuttal has been posted. The rebuttal period is a critical part of our review process, allowing authors to clarify misunderstandings and address specific concerns.

Please take a moment to read the author response, engage in the discussion by replying to the authors' responses, and update your review if the rebuttal changes your assessment.

Your active participation is vital to ensuring a fair and thorough evaluation of the submissions. A constructive discussion helps us make the most informed decisions.

Thank you for your time and valuable contributions to the review process.

Best regards,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.