acceptodds
Under review as a conference paper at ICLR 2027

ImpactFlow: Guiding Flow Matching with Geometry from Future Outcomes

Abstract

Generative models are trained through local prediction signals, yet their quality is judged on completed outputs. Under limited capacity and computation, how should learning prioritize local errors according to their future consequences? In flow matching, equal-norm velocity errors receive equal local cost even when the remaining dynamics carry them into different output changes. We introduce **terminal-transported residual geometry**, which relates error directions at the current state to their predicted effects after the remaining generation process. Output structure specifies which changes matter, and the remaining dynamics determine how local directions contribute to those changes. **ImpactFlow** instantiates this view as directional residual weighting for flow matching, preserving the original conditional velocity labels and the deployment interface. Experiments on low-step HumanML3D text-to-motion generation show that transported geometry predicts independently measured output changes better than decoder-only geometry. Controlled training comparisons demonstrate improved generation quality over endpoint geometry after matching scalar training strength, supporting directional future impact as an effective learning signal. These results support a learning principle: local errors should be prioritized by their future impact on the generated outcome.

Primary area:
generative models
Keywords:
Flow matching, generative modeling, terminal sensitivity, residual geometry, low-step generation, text-to-motion generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_wW8ZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ImpactFlow modifies flow-matching training to penalize prediction errors according to their effect on the final generated output. It estimates this sensitivity by differentiating through a frozen teacher’s remaining flow, decoder, and motion features. Random probes make the additional loss practical, and inference remains unchanged.

Experiments on HumanML3D evaluate continuation from a pretrained model. The paper compares against endpoint-based controls, studies the sensitivity estimates, and reports results under both matched-update and matched-time budgets.

Critical strengths
  • The method addresses a meaningful limitation of the standard FM objective. Equal-sized velocity errors can have different downstream effects. Accounting for the remaining flow gives a concrete way to distinguish these directions, while leaving inference unchanged. This is the main methodological contribution in my assessment.

  • The controls provide convincing evidence that the transported direction matters. The endpoint-matched comparison controls several scalar factors that could otherwise explain the improvement. ImpactFlow improves FID in nine of ten seeds, with a mean relative improvement of 13.23%. Together with the sensitivity and direction-shuffling experiments, this supports the proposed mechanism beyond a simple increase in auxiliary supervision.

Critical weaknesses
  • The practical benefit under a fixed compute budget remains uncertain. Each update requires additional teacher and sensitivity computation. In the matched-time experiment, the mean FID improvement is 10.87%, but its 95% confidence interval is [−1.86%, 23.60%]. The matched-update results support the value of the training signal, but the evidence that it pays for its computational cost is less conclusive. This limits the practical significance of the reported gains.

  • The evaluation does not establish whether the FID gains translate into broader motion-quality improvements. The method explicitly uses joint positions, velocities, bone lengths, and contact-related features, yet the reported generation results do not adequately cover physical quality, text alignment, or diversity. These measurements matter because improving FID alone does not rule out tradeoffs in other aspects of generation. The restriction to one dataset, model, and short continuation setting further limits the scope of the evidence.

Questions for authors
  • Clarification on compute efficiency: How sensitive is the matched-time result to the chosen continuation budget? If existing measurements cover other budgets, do they show a consistent advantage after including all geometry-computation costs?

  • Clarification on generation quality: Do you have text-alignment, diversity, or physical-quality measurements for the same checkpoints? These would materially affect my assessment of whether the FID improvement represents a broader quality gain.

3 comments
AuthorsOpenReview ↗

Official Comment by Authors

We thank you for the careful assessment of both the computational cost and the motion quality of ImpactFlow. Your questions motivated us to strengthen the end-to-end equal-time evaluation, examine multiple continuation budgets, and report text-alignment, diversity, multimodality, and foot-skating metrics on the same checkpoints as the main FID comparison. We also broadened the evaluation to KIT-ML and additional backbone sizes. We respond to each point below.

W1/Q1 — Practical benefit and sensitivity to continuation budget

We appreciate the focus on end-to-end compute, since the practical value of ImpactFlow should be assessed with geometry estimation included. We therefore expanded the equal-time HumanML3D comparison to 32 paired models. Both methods receive 1,919 seconds on four A100 GPUs; the ImpactFlow budget includes teacher rollout, sensitivity computation, student updates, startup, and finalization. Both methods use the same Standard-FM-selected midpoint NFE 8 / CFG 3 deployment.

MethodContinuation cap (s)Mean student updatesMean FID ↓
Standard FM1,91947,793.840.10206
ImpactFlow1,91914,130.130.09106

Across the 32 paired models, the absolute FID reduction is 0.01100 (95% CI [0.00215, 0.01985], exact two-sided p = 0.01707), and the mean modelwise relative reduction is 10.03% (95% CI [1.18%, 18.88%], p = 0.02809). Each model is evaluated four times, with the independently trained model as the statistical unit.

We also examined three continuation budgets using the same eight prespecified parent models. Each budget uses independently trained continuations, and geometry computation is included in the time cap.

BudgetCap (s)ModelsStandard FM FID ↓ImpactFlow FID ↓95% CI of reductionExact p
Half959.580.104860.08117[0.01099, 0.03639]0.00781
Standard1,919.080.104440.09308[-0.00365, 0.02637]0.11719
Double3,838.080.093730.08228[-0.00404, 0.02694]0.12500

ImpactFlow obtains lower mean FID at all three continuation budgets. This budget study complements the larger equal-time result and provides a direct view of how the method behaves as the available continuation time changes.

AuthorsOpenReview ↗

Official Comment by Authors

W2/Q2 — Motion quality on the same checkpoints and evaluation breadth

Thank you for emphasizing that FID should be considered together with text alignment, diversity, and physical motion quality. We measured these metrics on the same 32 pairs of HumanML3D checkpoints used for the main FID comparison.

MethodR-Precision @1 ↑MM-Dist ↓DiversityMultimodalityFoot-skating rate ↓
Standard FM0.44623.3309.4721.9777.24%
ImpactFlow0.44473.3259.4622.0377.51%
Holm-adjusted p0.289880.594090.594096.52 × 10⁻⁹0.00024

The table reports text alignment, overall diversity, multimodality, and foot skating on the same checkpoints as the primary FID evaluation. ImpactFlow also increases Multimodality from 1.977 to 2.037. This provides a direct, metric-by-metric account of how the methods compare beyond FID.

We further broadened the evaluation to KIT-ML and two additional HumanML3D backbone sizes:

StudySettingStandard FM FID ↓ImpactFlow FID ↓
KIT-ML16 models, 2,000 pretraining epochs, matched continuation updates, NFE 82.641032.60573
HumanML3D16.8M backbone, single-seed evaluation, NFE 80.2250.200
HumanML3D130M backbone, single-seed evaluation, NFE 80.1570.133

On KIT-ML, the mean modelwise relative FID reduction is 1.32%; the 95% CI for the absolute Standard-minus-ImpactFlow reduction is [0.00294, 0.06767], with exact paired p = 0.02304. These additions extend the evaluation across datasets, model sizes, training budgets, and complementary motion-quality measures.


Thank you again for emphasizing the importance of evaluating compute cost and motion quality together. The expanded equal-time study and the additional quality metrics provide a more complete account of ImpactFlow's behavior under the requested settings. We hope this revised evidence and presentation address your questions.

AuthorsOpenReview ↗

Follow-up on Additional Efficiency and Motion Quality Results

Dear Reviewer wW8Z,

Thank you again for your thoughtful review and for emphasizing the importance of evaluating computational efficiency and motion quality together. We particularly appreciate your suggestions regarding the equal-time setting, continuation-budget sensitivity, and complementary motion-quality metrics.

Following your suggestions, we expanded the equal-time HumanML3D study to 32 independently pretrained models, including all teacher rollout and sensitivity computation costs. The expanded study shows a 10.03% mean relative FID reduction under the same continuation-time budget. We also evaluated three continuation budgets and reported R-Precision, MM-Dist, Diversity, Multimodality, and foot-skating measurements on the same checkpoints used for the main FID comparison. In addition, we added KIT-ML experiments and evaluations with different backbone sizes.

We hope these additional experiments provide a clearer and more comprehensive picture of ImpactFlow's practical effectiveness and motion-generation quality.

As the discussion deadline is approaching, we wanted to kindly ask whether you have had a chance to review our updated responses. We would greatly appreciate any further thoughts on whether the additional evidence addresses your questions.

Thank you again for your time, thoughtful feedback, and consideration.

Reviewer_YTryOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper trains a flow matching model by prioritizing local errors that change the final output by the greatest degree. The authors call their training method ImpactFlow that uses "terminal transported residual geometry" and perform experiments on low-step HumanML3D text to motion generation.

Critical strengths

S1. The method is novel and well motivated, and demonstrates improvement during sampling with minor changes in training structure.

S2. Some experiments are sound, using fresh seed replication and reports on non significant results, demonstrating rigour in analysis.

Critical weaknesses

W1. Some key baselines are missing. Simple scalar weighting and timestep reweighting for example are missing. There is also no eval on non-low-step pipelines. Some of the controls that are defined in the paper are also not reported. While some experiments demonstrate rigour, there are some that have gaps (eg. FID results from controls like Diagonal, shuffle)

W2. While the idea is novel, reported results are quite weak with only FID as a metric (on one dataset). At proper setting the gain is not significant.

W3. The least accurate estimator works best which do not clearly justify that the gain is coming from geometry or regularisation. It would be good if the authors can clarify this.

Questions for authors

See weakness. I will raise my score if these are resolved.

4 comments
AuthorsOpenReview ↗

Official Comment by Authors

We thank you for the thoughtful review and, in particular, for asking whether the gains come from directional geometry rather than from a simpler reweighting effect. In the revision, we added sample-level and timestep-level scalar controls, a randomized-direction control under an equal-time budget, and the Diagonal and shuffled-direction results. We also expanded the evaluation to KIT-ML and additional motion-quality metrics, and further analyzed the estimator's accuracy–compute trade-off. Our point-by-point responses follow.

W1 — Scalar weighting, directional controls, and many-step sampling

We appreciate the suggestion to compare against simple scalar weighting and randomized directions. These controls are particularly useful for identifying whether the benefit comes from the geometry's orientation. We added an equal-time comparison on ten paired HumanML3D models at Euler NFE 8 / CFG 3. All methods use the same continuation-time budget, including teacher rollout and terminal VJP computation.

ControlControl FID ↓ImpactFlow FID ↓95% CI of paired reductionHolm p
One weight per sample0.331250.25612[0.05120, 0.09906]0.00586
One weight per timestep0.329700.25612[0.04966, 0.09750]0.00586
Random direction, original vector length0.334810.25612[0.05616, 0.10122]0.00586

ImpactFlow outperforms all three controls after Holm correction. The sample- and timestep-level controls assign a common weight to all residual directions within a sample or timestep. The randomized-direction control preserves weighting strength while changing the orientation. Together, these comparisons show that the directional structure estimated by ImpactFlow contributes beyond scalar reweighting.

We also include the Diagonal and shuffled-direction controls from the five-seed study. Their mean FIDs are 0.246 and 0.235, respectively, compared with 0.222 for ImpactFlow (Holm-adjusted p = 0.375 for each comparison). These results are reported in Appendix I.6.

We additionally report the high-NFE evaluation in Appendix I.6 (Heun NFE 1024), where Standard FM and ImpactFlow achieve FIDs of 0.091 and 0.137, respectively. Our primary focus is low-step generation, where the matched-budget experiments at NFE 4 and 8 provide the strongest evidence for ImpactFlow's effectiveness.

AuthorsOpenReview ↗

Official Comment by Authors

W2 — Statistical significance, motion-quality metrics, and KIT-ML

We agree that the evaluation should include statistical evidence, complementary motion-quality metrics, and another dataset. We therefore expanded the HumanML3D equal-time study to 32 paired models and added a 16-model KIT-ML comparison with matched continuation updates.

Dataset / matchingModelsStandard FM FID ↓ImpactFlow FID ↓95% CI of Standard-minus-ImpactFlowExact paired p
HumanML3D / time320.102060.09106[0.00215, 0.01985]0.01707
KIT-ML / updates162.641032.60573[0.00294, 0.06767]0.02304

The mean modelwise relative FID reductions are 10.03% on HumanML3D and 1.32% on KIT-ML. Both comparisons use paired models, with the reported confidence intervals and exact paired tests computed at the model level.

For the same 32 HumanML3D model pairs used in the equal-time FID comparison, we report the following complementary metrics:

MetricStandard FMImpactFlowHolm-adjusted p
R-Precision @1 ↑0.44620.44470.28988
R-Precision @2 ↑0.64910.64940.71105
R-Precision @3 ↑0.75810.75880.63940
MM-Dist ↓3.3303.3250.59409
Diversity9.4729.4620.59409
Multimodality1.9772.0376.52 × 10⁻⁹

These results provide a broader assessment of text alignment and motion diversity on exactly the same checkpoints as the primary FID comparison. In particular, ImpactFlow improves Multimodality while maintaining comparable R-Precision, MM-Dist, and Diversity.

AuthorsOpenReview ↗

Official Comment by Authors

W3 — Why the inexpensive estimator can work well

This question gets at an important design choice: under a fixed compute budget, estimator accuracy must be considered together with the number of student updates it permits. We therefore report both a fixed-checkpoint gradient diagnostic and an equal-time continuation study.

First, to isolate the effect of probe averaging, we use the same two-step teacher map for the estimate and the 256-probe reference:

Gradient estimateCosine to the same-depth reference ↑
Single independent estimate0.484
Mean of 16 independent estimates0.921

Averaging substantially improves gradient agreement. We then vary teacher depth and probe count under the same continuation-time cap. The following results are means over three training seeds; gradient cosine is measured against a common eight-step, 256-probe reference.

Teacher Heun stepsProbesGradient cosine ↑Student updatesEuler FID ↓Midpoint FID ↓
210.3349,311.70.211480.08450
240.5483,912.70.210130.08727
810.5742,856.30.278890.12625
840.8321,082.70.238200.09272

The one-probe, two-step estimator gives the student the largest number of updates within the fixed budget and achieves the lowest observed mean FID with the midpoint sampler. The randomized-direction experiment in W1 further shows that the direction estimated by ImpactFlow is useful for training. Taken together, these experiments explain why the lightweight estimator is an effective choice in our compute-constrained setting.


We appreciate your suggestions on the baselines and estimator analysis. The added scalar and randomized-direction controls directly examine the role of directional information, while the expanded evaluation places the results in a broader generation setting. We hope these additions address your concerns and clarify where ImpactFlow provides its strongest benefits.

AuthorsOpenReview ↗

Follow-up on Additional Controlled Experiments

Dear Reviewer YTry,

Thank you again for your constructive suggestions, particularly regarding scalar weighting baselines, directional controls, and the accuracy of our geometry estimator.

In response, we added equal-time comparisons against sample-level weighting, timestep-level weighting, and a randomized-direction control. ImpactFlow achieves significantly lower FID than all three controls after Holm correction. We also reported the Diagonal and shuffled-direction experiments, expanded the evaluation to KIT-ML and additional motion-quality metrics, and included a many-step sampling evaluation.

Regarding the estimator, we further examined gradient agreement under matched teacher depth and analyzed the trade-off between estimator accuracy and the number of student updates under a fixed compute budget.

We believe these additional experiments provide direct evidence for the role of directional geometry in ImpactFlow and address the main concerns raised in your review.

With the discussion deadline approaching, we would be grateful if you could take a look at our updated responses and let us know whether the additional evidence addresses your concerns. We would be happy to clarify any remaining questions.

Thank you again for your time and helpful feedback.

Reviewer_VLAnOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes ImpactFlow, which weights flow-matching errors by how much they would change the final output. A frozen teacher rolls the current state forward, and one VJP estimates which error directions matter most. This weight is used only in training, so the model and sampler stay the same. On HumanML3D text-to-motion with 8 steps, it lowers FID compared to Standard FM and to endpoint-only geometry.

Critical strengths

1.Clear and well-motivated idea. Equal-size errors can have very different effects on the final output. The paper turns this into a training weight that keeps the original FM target and adds no cost at sampling time. This is a sound and easy-to-use idea.

2.Careful control experiments. The scalar-matched controls (Token-sign, Token-matched, Endpoint-matched) show that the direction of the weight matters, not just its size. The diagnostic in Figure 3 also shows that transported geometry predicts output changes better than decoder-only geometry. This makes the gains more believable than a simple loss change.

Critical weaknesses

1.Very narrow tests. There is only one dataset (HumanML3D), one small model (39M), one sampling setting (8 NFE), and one metric (FID). The paper claims a general learning principle, but shows it in only one setting, so we cannot tell if it holds elsewhere.

2.Missing standard metrics. R-Precision, MM-Dist, and Diversity are not reported. These are standard for text-to-motion. A lower FID may come with worse text match or diversity, and we cannot check this.

3.No strong baselines. The paper only compares to Standard FM, LPL-style, and its own controls. It does not compare to few-step methods (consistency models, shortcut models, MeanFlow), which it cites as the main motivation, or to known text-to-motion models. So we cannot judge how useful it is in practice.

4.Weak practical result. In the equal-time test (Table 2), the 95% CI is [−1.86, 23.60] and includes zero, with only 7 of 10 seeds favorable. Most other tests use only 5 seeds, and the fixed-update tests ignore the extra cost of the teacher rollout and VJP. The real-world gain is not clearly shown.

5.Noisy estimate. The default estimator has gradient cosine of only 0.334 with the reference, and it wins mainly because it is cheap. It is unclear how much of the gain comes from the geometry itself rather than from noisy reweighting. Hard to read. The paper uses many custom terms and long, dense sentences. A simple idea becomes hard to follow, which hurts clarity and impact.

Questions for authors

1.Can you report R-Precision, MM-Dist, and Diversity for the main results, and add a second dataset such as KIT-ML? This is a clarification, and it is key to my score. 2.How does ImpactFlow compare to few-step methods such as consistency models, shortcut models, or MeanFlow at the same NFE and compute? This is a clarification. 3.Can you run more seeds for the equal-time test (Table 2) to show whether the gain is significant? This is a clarification.

5 comments
AuthorsOpenReview ↗

Official Comment by Authors

We thank you for the careful review and for highlighting the need to strengthen the empirical evaluation of ImpactFlow. Your questions about evaluation breadth, few-step baselines, equal-time efficiency, and estimator reliability helped us identify the most important additions for the revision. We have added KIT-ML results and standard motion-quality metrics, compared against Shortcut under a matched continuation-time budget, expanded the equal-time HumanML3D study to 32 independently pretrained models, and added a direct randomized-direction control. We address each point below.

W1/W2/Q1 — Evaluation breadth, text alignment, and motion diversity

We appreciate this suggestion. Evaluating across datasets and complementary metrics is important for understanding whether the observed FID improvement is accompanied by consistent text-to-motion quality. We therefore added KIT-ML and report R-Precision, MM-Dist, and Diversity alongside FID.

The NFE 8 results are:

DatasetMethodFID ↓R-Precision @1 / @2 / @3 ↑MM-Dist ↓Diversity
HumanML3DStandard FM0.102060.4462 / 0.6491 / 0.75813.3309.472
HumanML3DImpactFlow0.091060.4447 / 0.6494 / 0.75883.3259.462
KIT-MLStandard FM2.641030.3922 / 0.5970 / 0.71513.29910.765
KIT-MLImpactFlow2.605730.3944 / 0.5984 / 0.71743.29310.766

The HumanML3D comparison uses 32 independently pretrained models under matched continuation time. The KIT-ML comparison uses 16 models pretrained for 2,000 epochs under matched continuation updates. On KIT-ML, ImpactFlow reduces mean modelwise FID by 1.32% (paired p = 0.02304). The complementary metrics provide a broader view of generation quality on both datasets.

We also evaluated ImpactFlow with two additional HumanML3D backbone sizes:

Backbone sizeStandard FM FID ↓ImpactFlow FID ↓
16.8M0.2250.200
130M0.1570.133

ImpactFlow obtains lower FID in both model-size evaluations. These results, together with the KIT-ML study, extend the empirical evaluation beyond the original HumanML3D setting.

AuthorsOpenReview ↗

Official Comment by Authors

W3/Q2 — Few-step baselines and existing text-to-motion models

This is an important comparison for assessing the practical value of ImpactFlow in the low-step regime. We added Shortcut, a method specifically designed for few-step generation, to a study with 16 independently pretrained HumanML3D models. Standard FM, Shortcut, and ImpactFlow use the same backbone and pretrained starting points, with the same 1,919-second continuation-time budget. Method-specific training computation is included in the budget. Validation selects midpoint / CFG 3 for Standard FM and ImpactFlow, and native Euler / CFG 4 for Shortcut.

NFEStandard FM FID ↓Shortcut FID ↓ImpactFlow FID ↓Holm p: vs Standard FMHolm p: vs Shortcut
40.269560.319230.241430.002440.00073
80.097260.138860.089630.152830.00009
160.075830.079940.077181.000001.00000

At NFE 4, ImpactFlow achieves a 10.43% mean paired relative FID reduction over Standard FM. At NFE 8, it achieves a 34.55% reduction over Shortcut. The results demonstrate the benefit of ImpactFlow in the few-step generation setting that motivates our method.

For additional context, we evaluated released MARDM and MoRAE models using the same motion evaluator and their respective sampling procedures:

Released modelSampling settingFID ↓
MARDM8 autoregressive iterations0.184
MoRAE7 field evaluations0.116

These evaluations complement the controlled comparison with Shortcut and provide additional reference points against established text-to-motion systems.

AuthorsOpenReview ↗

Official Comment by Authors

W4/Q3 — Equal-time efficiency and statistical evidence

We agree that the additional teacher and sensitivity computation must be included when evaluating practical efficiency. To address this directly, we expanded the equal-time HumanML3D study to 32 independently pretrained models. Each continuation receives 1,919 seconds on four A100 GPUs, including teacher rollout, sensitivity computation, student updates, startup, and finalization. Both methods are evaluated at the same midpoint NFE 8 / CFG 3 deployment, selected using Standard FM.

MethodContinuation cap (s)Mean student updatesMean FID ↓
Standard FM1,91947,793.840.10206
ImpactFlow1,91914,130.130.09106

Across the 32 paired models, the absolute FID reduction is 0.01100 (95% CI [0.00215, 0.01985], exact two-sided p = 0.01707). The mean modelwise relative reduction is 10.03% (95% CI [1.18%, 18.88%], p = 0.02809). Each model is evaluated four times, and the independently trained model is the statistical unit.

This expanded study directly demonstrates that ImpactFlow improves FID under an equal continuation-time budget that includes the full geometry-computation cost.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Author Statement: ImpactFlow's Motivation and Contributions (Submission 5700)

Dear AC, SAC, PCs, and Reviewers,

We sincerely thank all reviewers for their thoughtful feedback and the AC, SAC, and PCs for their time and effort throughout the review process. In response, we have strengthened our experiments and clarified the motivation and contributions of our work. We would like to highlight the central idea of ImpactFlow and the key evidence from our revision.

1. Motivation and Method.

Our work addresses a fundamental question in generative learning: When training resources are limited, which local prediction errors should a model prioritize to improve its final outputs?

In standard Flow Matching, velocity errors of equal magnitude receive the same regression penalty, even though the remaining generation dynamics can propagate them into very different final outputs. ImpactFlow addresses this by propagating output sensitivity backward through the remaining flow of a frozen reference generator, assigning different weights to residual directions according to their predicted downstream effects. Importantly, it preserves the original conditional Flow Matching targets and standard inference procedure.

2. Addressing the Reviewers' Main Concerns.

The reviewers' main concerns focused on practical effectiveness, the contribution of directional geometry, and evaluation breadth. We conducted additional experiments to address these questions.

  • Equal-time effectiveness. We expanded the HumanML3D study to 32 independently pretrained models, including all teacher rollout and sensitivity computation costs. ImpactFlow reduces mean FID from 0.10206 to 0.09106, achieving a 10.03% mean relative reduction (95% CI: [1.18%, 18.88%], p = 0.01707).

  • Few-step baselines and broader evaluation. At NFE = 8, ImpactFlow achieves an FID of 0.08963, compared with 0.13886 for Shortcut, a 34.55% mean paired relative reduction (Holm-adjusted p = 0.00009). We also added KIT-ML, different backbone sizes and sampling budgets, and evaluations of released MARDM and MoRAE models.

  • Directional geometry. We added sample-level, timestep-level, and randomized-direction controls under matched continuation-time budgets. ImpactFlow significantly outperforms all three (Holm-adjusted p = 0.00586 for each comparison). The randomized-direction control preserves weighting strength and computational cost while changing only the direction, directly supporting the contribution of directional geometry.

  • Estimator and motion quality. Averaging 16 independent estimates increases gradient cosine similarity from 0.484 to 0.921 under the same teacher map. We also report R-Precision, MM-Dist, Diversity, Multimodality, and foot-skating measurements on the same checkpoints as the main FID comparison.

3. Key Findings and Contributions.

Our experiments highlight two findings. First, the remaining generation dynamics contain useful information about which local error directions matter most: transported geometry better predicts independently measured output changes than decoder-only geometry, and controlled experiments demonstrate the value of this directional information.

Second, estimation accuracy and learning progress should be considered jointly. A lightweight estimator can provide useful directional guidance while leaving more computation for student updates under a fixed budget.

Together, these findings establish a learning perspective in which local prediction errors are prioritized according to their future impact on generated outcomes, rather than their magnitude alone.

4. Closing Remarks.

We are grateful for the reviewers' suggestions, which helped us strengthen the empirical evidence and clarify the underlying mechanism. We sincerely hope the final assessment will consider the motivation, methodological contribution, controlled evidence, and expanded evaluation of ImpactFlow as a whole.

Thank you again for your time and consideration.

Best regards,

The Authors of Submission 5700

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.