acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Prediction Horizons: Decoupling Model Queries from State Updates in Few-Step MeanFlow

Abstract

MeanFlow predicts average velocity over a queried time interval, while standard few-step sampling ties this prediction horizon to the state-update duration. We decouple these quantities at inference, querying a longer horizon while keeping the state grid, update durations, checkpoint, and query count fixed. Across seven independently trained CIFAR-10 checkpoints at classifier-free guidance (CFG) scale 2, longer horizons worsen FID early in sampling but improve it at the midpoint. This benefit also depends on the incoming-state distribution. In held-out factorial outcomes on five previously studied checkpoints, with a hypothesis frozen before inspection, changing only preceding CFG while fixing current and subsequent guidance increases the mean horizon benefit from 0.109 to 1.308 FID. A frozen four-query schedule reduces mean CIFAR-10 FID by 2.023 without retraining or additional queries, including improvements on five checkpoints trained after rule selection. A milder frozen rule further improves mean FID by 1.041 on a stronger grid selected from 84 predefined timestep candidates and transfers without retuning to the evaluated public ImageNet-256 and AFHQ checkpoints. Under matched new calibration budgets, start-time shifting achieves lower FID/KID than horizon-only adjustment in historical CIFAR tests, while the tested Horizon allocation adds no FID benefit beyond Grid+Start on two public ImageNet models. These results identify prediction horizon as one of several useful temporal controls: the physically matched interval need not yield the best endpoint metric. Prediction horizon is a configuration-dependent inference control with no additional queries after selection; the evaluated selection procedures use target rollouts.

Primary area:
generative models
Keywords:
MeanFlow, flow matching, few-step generation, prediction horizon, inference-time calibration

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_2QJxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Building upon the foundation of established MeanFlow, which conventionally treats the prediction horizon and the state-update duration as equivalent, therefore this paper reveals that the distance of model prediction and the distance of actual movement do not necessarily need to be simply equated. Next, it demonstrates the necessity of distinguishing between the prediction horizon and the step size. Furthermore, the study also finds that extending the prediction horizon does not consistently yield performance improvements in terms of FID/KID across different query positions. Specifically, longer horizons worsen performance at earlier sampling positions, such as t=0t=0 and t=0.25t=0.25, while improving performance around the midpoint t=0.5t=0.5, a phenomenon characterized as the midpoint effect. In the experiments, the authors compare the horizon strategy with other temporal control methods under constrained query budgets, showing that the relative effectiveness of different temporal controls depends on the tested models, query budgets, and evaluation metrics.

Critical strengths

To be honest, I am not deeply familiar with the specific technical details of the MeanFlow work. Nevertheless, I can clearly grasp the significance of the authors' motivation. They rethink the image generation process by decoupling the prediction horizon from the state-update duration, in particular considering how far the model should look (predict) ahead while keeping the actual movement of the sampling process fixed. Compared with the conventional MeanFlow setting that directly couples these two quantities, this research explores whether adjusting the prediction horizon can lead to better generation quality without increasing the number of model queries. Furthermore, both the paper's experimental validation provide solid support for the authors' hypotheses and final conclusions.

Critical weaknesses

None. The authors have already acknowledged the limitations of horizon selection in Section 6 (Discussion), which directly corresponds my main concern. But I have some questions below.

Questions for authors

Although Figure 4 presents a qualitative comparison between the Local and Horizon methods, the visual differences are relatively subtle. Therefore, for the sake of completeness, I would like the authors to provide a more detailed qualitative analysis or representative examples of failure and success cases. Additionally, could the authors clarify specifically which aspects of perception saw improvements or degradations due to the Horizon control?

Another important point concerns that why the benefit of horizon extension changes from degradation at t=0.25t=0.25 to improvement at t=0.5t=0.5. Although the authors acknowledge the possible sources of this effect in the discussion section, I would like them to further elaborate on the underlying attribution of this benefit. It would provide more compelling evidence if the authors could conduct a quantitative results between these query positions to examine how the horizon performance grows, decays, or changes sign.

Could the authors maintain a consistent placement of table captions throughout the paper?

3 comments

Paired examples and distributional measurements

Thank you for requesting clearer examples and a more specific account of what improves. Revised Figure 8 in Appendix E presents six paired Local/Horizon/direction-only examples. Within each triplet, checkpoint, initial noise, requested class and sample index are identical.

The six cases come from a 150-case pool of Local/early-Horizon/midpoint-Horizon triplets. Figure 8 adds existing direction-only outputs at six of those indices. The supplementary files contain the full original pool and contact sheets; the figure uses saved evaluation pixels without enhancement. This selection provides illustrative cases rather than a prevalence estimate.

In panels c and e, both midpoint Horizon and direction-only outputs appear clearer than Local. Panel c remains semantically ambiguous even after this change. The informal inspection found no visually discernible difference between Horizon and direction-only in these examples. These examples illustrate clearer appearance alongside persistent semantic ambiguity. They were assessed informally, rather than in a human preference study.

The distributional evaluations show a measurable tradeoff. In the separate seven-checkpoint classifier-free guidance (CFG) 1.5 comparison on identical saved images, Horizon lowers DINOv2 feature-space precision from .7171 to .6967 and raises recall from .3617 to .3921 relative to Local. DINOv2 distance, Inception FID and KID all improve. In this representation, more real features fall within the generated-feature neighborhoods, while fewer generated features fall within the real-feature neighborhoods. It is not a direct perceptual rating of Figure 8, which belongs to a different experiment.

The calibrated four-step comparison has a different pattern: Combination improves Inception precision and recall over Scalar, from .7369/.4481 to .7433/.4594, while requested-class agreement falls slightly from 98.44% to 98.29%. Appendix C reports those measurements, and Appendix E separates the evaluator-specific results.

Attribution of the position contrast

We added direct interventions to examine why midpoint adjustments produce larger benefits. We also refined the premise concerning t=.25t=.25: in the updated seven-checkpoint scan, h=.50h=.50 gives a mean gain of +.105 FID with 6/7 improvements, whereas h=.75h=.75 worsens all seven. Earlier small negative quarter-point results remain reported under their respective protocols. The consistent source-to-midpoint reversal and the quarter-to-midpoint benefit difference are treated separately.

For the quarter-to-midpoint comparison, RefLocal and RefLong are average velocities obtained by integrating the model's guided instantaneous field over the local and long prediction intervals. We apply each with the same physical update duration, δ=.25\delta=.25. Their comparison removes the Local predictor's discrepancy from this integrated model reference at the changed query.

We then hold each arm's t=.75t=.75 state fixed across final samplers and change only the interval [.75,1]. For the t=.25t=.25 intervention, propagation through [.50,.75] remains native. Let DD be the RefLong-over-RefLocal FID gain at t=.50t=.50 minus that at t=.25t=.25. On four confirmation checkpoints, mean DD is 3.421 with one native final update, 2.699 with two native half-steps and .073 with fourth-order Runge–Kutta (RK4) integration. Every checkpoint has this ordering in FID and KID. The one-native-step minus RK4 difference is 3.348 FID, with descriptive 95% interval [2.894, 3.803] (Appendix D.7).

Replacing only the final interval is sufficient to substantially reduce the measured position contrast. The interval for the smaller RK4 residual spans zero, leaving the residual contrast uncertain. These final-interval numbers also differ from the separate experiment that replaces the entire remaining suffix in Appendix D.6.

At the midpoint with full-amplitude reference interventions, the three exploration checkpoints provide additional distributional detail. Native and RK4 final maps yield similar feature-mean displacement lengths, 1.156 and 1.173, but mean-term FID gains of 1.347 and .142 and covariance-term gains of 2.133 and .182. Writing the baseline mean error as b=μreal−μbaseb = \mu_{\mathrm{real}} - \mu_{\mathrm{base}} and the intervention's mean shift as dd, the mean-term gain is

Gμ=2b⊤d−∥d∥2.G_{\mu} = 2b^{\top}d - \lVert d \rVert^2. A displacement of similar length can therefore have a different benefit when the baseline error and its alignment with the displacement change. The final sampler changes both quantities; the covariance term also contributes to the observed FID difference.

This is a more specific explanation than attributing the effect to a larger velocity change: the correction interacts with the subsequent sampling map and the baseline distribution it produces. The reference field comes from the learned model, and RK4 uses extra computation for diagnosis. The specific network-error source remains unresolved (Section 4; Appendices D.5–D.7).

Intermediate-position measurements and table captions

Figure 7 adds the requested intermediate-position measurements at s=.30s=.30, .35, .40, .45 and .50. Each uses the grid (0,.25,s,s+.25,1)(0,.25,s,s+.25,1); the third query has physical update duration .25 and prediction horizon .50. Local and Horizon share the entering state before that query and paired generation inputs.

Position s.30.35.40.45.50
Mean FID gain.243.6301.0511.4942.002
Mean KID gain x1000.402.646.9081.1731.470

Positive gain favors Horizon. Both FID and KID gains increase across the measured positions on each of the three checkpoints, and all measured positions have positive mean gains. Figure 7 shows the individual checkpoints as well as their mean; the supplementary files contain the paired numerical results (Appendix D.12).

The scan establishes an increasing benefit within this grid family. Moving s also changes the prefix and suffix durations, and the original t=.25t=.25 intervention occurs at a different query index. Consequently, we do not join it to this series as a pure time-position sign-change experiment. The fixed-state final-interval interventions in Section 4 and Appendix D.7 test downstream sampling dependence separately.

We also standardized table captions above the tables throughout the manuscript and appendices. Figure numbers in these responses refer to the revised manuscript: the horizon-response curves are Figure 6, the intermediate-position scan is Figure 7, and the qualitative examples are Figure 8.

Dear authors,

Thanks for your adequate rebuttal with my concerns.

Proposed suggestions are anticipated to incorporate into the revised paper and imporve the clarification of paper.

So, I decide to remain the exisiting postitive score.

Reviewer_1gSeOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

MeanFlow predicts an average velocity over a requested time interval, and standard sampling uses the same interval for the state update. The authors decouple the prediction horizon from the update duration while keeping the model, sampling grid, and query count fixed. In guided CIFAR-10 experiments, extending the horizon at the midpoint improves FID, whereas applying longer horizons at earlier sampling positions worsens mean FID. A milder rule is also evaluated on ImageNet and AFHQ.

Critical strengths
  • From a methodological perspective, the paper studies an interesting inference-time choice and carefully isolates changes in prediction horizon from changes in update duration.

  • The proposed midpoint horizon extension is validated across seven independently trained CIFAR-10 checkpoints, achieving a 2.0232.023 FID improvement without retraining or additional model queries. A milder adjustment also achieves a 1.0411.041 FID improvement on a stronger sampling grid.

Critical weaknesses
  • The motivation is not convincing. The paper does not adequately justify why deliberately mismatching the prediction horizon and update duration should improve generation. Although the formulation distinguishes horizon adjustment from changing the update coefficient δk\delta_k, the authors do not give a compelling intuition for why this is a scientific problem that is worth studying rather than simply an empirically tuned perturbation.

  • The paper does not explain its central empirical finding. Although the authors provide an error decomposition and exploratory velocity diagnostics, these do not explain why longer horizons improve FID at the midpoint but worsen it earlier in sampling. As the position-dependent reversal is the paper's main finding, the absence of a convincing explanation substantially weakens its scientific contribution.

  • The proposed midpoint adjustment is ad hoc. The horizon rule is empirically selected, with no established principle for predicting which sampling position or horizon will improve generation.

  • Minor issue: The distinction between “Horizon” as the intervention name and “horizon” as the prediction duration should be made consistent.

Questions for authors
  • Can the authors provide plots showing how the predicted velocity's magnitude and direction change as the horizon varies at early and midpoint states, together with the corresponding FID changes? Such analysis is important for explaining the paper's central observation rather than relying primarily on endpoint FID improvements
  • MeanFlow allows one-step generation. However, the proposed model shows that expanding the horizon leads to worse generation at the starting point and requires applying the expanding horizon technique at the midpoint instead. Given this, I don’t see why we need to give up one-step generation to use the expanding horizon technique.
7 comments
AuthorsOpenReview ↗

Motivation, selection and terminology

1.Why distinguish horizon adjustment from scalar tuning?

Thank you for asking why a mismatch between prediction horizon and update duration is scientifically meaningful. For an exact interval-average predictor at unit velocity scale, matching the requested interval to the update duration gives the intended displacement. The question is what happens with a learned interval-conditioned predictor when quality is measured at the generated endpoint.

Section 2 makes the distinction explicit. The update is xnext=x+aδ,uθ(x,t,h).x_{\mathrm{next}} = x + a\delta\\,u_{\theta}(x,t,h).. Changing delta changes the physical step; changing a rescales the returned velocity; changing h requests a different prediction from the network. At a shared state, that prediction can change direction as well as magnitude. Subsequent updates then transform the resulting state displacement. This motivates testing whether the matched interval also gives the best endpoint distribution for the learned model.

We tested the most immediate alternative explanation: the benefit might simply reflect a useful velocity scale. In a separate one-step study, each classifier-free guidance (CFG) setting receives a 21-candidate scalar calibration on three development checkpoints and 2,000 shared inputs. Scalar retains the native one-step horizon h=1 during calibration, selecting a=.99 at CFG 1 and a=.98 at CFG 2. Combination then uses the selected scale with h=1.10; that horizon is held fixed in this study. On the four remaining checkpoints with new paired 10k inputs, Combination improves over Scalar by 1.523 and 1.362 FID, respectively; all four improve in FID and KID at each CFG (Section 3.4; Appendix D.2).

We then test direction matching at the calibrated CFG-2 scale. A within-class permutation projects a donor correction into the space orthogonal to the recipient's original velocity, then uses that direction while preserving the recipient's correction length, signed parallel component and angle. It also preserves the final velocity's norm and angle to the original velocity. Combination outperforms this control by 3.230 FID on average across four checkpoints, with descriptive 95% interval [1.295, 5.166]; all four favor Combination in FID and KID. Metrics are computed separately for each permutation bank, then averaged within each checkpoint before summarizing models.

Together, these comparisons show an increment beyond the tested calibrated scales and a dependence on which sample receives which direction. This supports studying horizon as a distinct control, while leaving the selection of a useful horizon as an empirical calibration problem. The four checkpoints were excluded from this scalar selection and had appeared in earlier studies. The direction comparison uses the same inputs as the scale confirmation.

2.Empirical selection, prediction and terminology

The midpoint rule remains empirically selected. We have made its role clearer: it is a tested operating point, while the broader contribution concerns what horizon changes do and when their benefit persists. Appendix C distinguishes inherited fixed-rule transfer from calibration on a target model or grid.

The revised evidence supports calibrating velocity scale and then testing the incremental value of horizon adjustment on separate generation inputs. It also shows why a single preferred horizon should not be assumed across settings. The one-step CIFAR example uses h=1.10, released MeanFlow benefits from h=.98, and the original native-scale AlphaFlow sweep prefers h=1. AlphaFlow gains from h=1.02 only in the subsequently calibrated-scale comparison, with a reduction in Inception Score (IS).

We investigated whether diagnostics could replace endpoint-based selection. Local feature-sensitivity forecasts are more reliable at the midpoint than at the source, and their selected mean gain does not exceed the fixed-position comparison. The velocity-disagreement predictor has a larger mean absolute error than the simpler checkpoint baseline (.303 versus .243), although its decision regret is lower (.00767 versus .01572). Appendix F reports these outcomes. The current analyses explain dependencies of the response, but do not provide a general predictor of the optimal horizon.

For terminology, we use “Horizon” for the named intervention and “horizon” for the prediction duration. The main deployment comparisons retain the original number of velocity queries. Magnitude/direction controls and numerical reference integrations are identified separately as additional-compute diagnostics.

AuthorsOpenReview ↗

Explaining the position-dependent effect

Direction controls and final-interval interventions

We now separate two observations that the earlier presentation grouped together: the source-to-midpoint sign reversal and the quarter-to-midpoint difference in benefit. Throughout this response, gain means baseline FID minus intervention FID, so positive values indicate improvement.

The updated seven-checkpoint scan shows that the quarter-point response is not uniformly negative. At t=.25t=.25, h=.50h=.50 gives a mean gain of +.105 FID, with 6/7 checkpoints improving; h=.75h=.75 worsens all seven. Earlier small negative results remain reported with their own input and evaluation settings. The more consistent sign reversal in the tested four-step setting is between t=0t=0 and t=.50t=.50 (Figure 1).

For that source-to-midpoint reversal, we first separate magnitude from direction. With update duration .25 and extended horizon .50, direction-only changes give mean FID gains over Local of -1.368 at the source and +1.818 at the midpoint, with the respective signs on all seven checkpoints. Next, we compare RefLong with RefLocal: interval averages constructed by integrating the model's guided instantaneous field over the long and local intervals. The FID gain of RefLong over RefLocal is -1.711 at the source and +3.655 at the midpoint. The reversal therefore persists after removing the local predictor's discrepancy from this model reference at the changed query (Section 4; Figure 3; Appendix D.5).

For the quarter-to-midpoint contrast, we test how the remaining sampler transforms the intervention. Let Gm(t)G_m(t) be the RefLong gain over RefLocal at position tt under final sampler mm, and define Dm=Gm(.50)−Gm(.25)D_m = G_m(.50) - G_m(.25). Each arm's state at t=.75t=.75 is held fixed across final samplers. On four confirmation checkpoints, with RK4 denoting fourth-order Runge–Kutta integration:

Final interval [.75,1]Mean D (FID)
One native update3.421
Two native half-steps2.699
RK4 on the model's h=0h=0 field0.073

The native-minus-RK4 interaction is 3.348 FID, with descriptive 95% checkpoint interval [2.894, 3.803]. Every checkpoint follows the same ordering in FID and KID. Earlier propagation remains native: this is a final-interval replacement, not replacement of the entire suffix (Appendix D.7).

At the midpoint with full-amplitude reference interventions, the three exploration checkpoints provide further distributional detail. Native and RK4 final samplers produce similar feature-mean displacement lengths, 1.156 and 1.173, but mean-term FID gains of 1.347 and .142. Covariance-term gains also differ, at 2.133 and .182. The final sampler changes both the response to the correction and the baseline distribution relative to the real data. Furthermore, increasing the correction from half to full amplitude increases the native-sampler benefit but reduces the RK4 benefit. The endpoint benefit therefore depends on correction amplitude as well as on the final sampler.

The evidence identifies direction dependence and a substantial final-sampler interaction. It does not yet locate a unique network-error source, and the quarter-to-midpoint controls are not a complete explanation of the distinct source-to-midpoint reversal. Reference integrations use the learned model field and additional computation (Section 4; Appendices D.5–D.7).

AuthorsOpenReview ↗

Requested plots and one-step generation

1.Velocity-response plots and endpoint quality

We added the requested plots in Figure 6, with the numerical data supplied in the supplementary files. At t=0t=0, .25 and .50 on the uniform four-query grid, the figure plots four quantities against prediction horizon: the guided-velocity norm ratio, its angle relative to the local-horizon prediction, FID gain and KID gain. All three diagnostic checkpoints are shown, together with their equal-weight mean.

Geometry is measured on 1,024 shared entering states per checkpoint and position. Endpoint quality uses separate paired 10k generation inputs. The panels compare geometry and endpoint quality for the same configurations using these separate input banks (Appendix D.12).

These curves help discriminate a simple geometric explanation. At h=.50h=.50, mean angles at t=.25t=.25 and t=.50t=.50 are similar, 4.29 and 4.51 degrees, yet the endpoint effects differ. Direction-only interventions further show opposite FID responses at source and midpoint. The change in direction matters, but the angle alone does not predict its benefit. Section 4 and Appendix D.7 examine this dependence by changing the final sampler while holding each intervention arm's entering state fixed.

Figure 7 adds positions s=.30s=.30, .35, .40, .45 and .50, using the grid (0,.25,s,s+.25,1)(0,.25,s,s+.25,1) with the intervention's update duration fixed to .25 and horizon to .50. Mean FID gains are .243, .630, 1.051, 1.494 and 2.002; KID gains multiplied by 1,000 are .402, .646, .908, 1.173 and 1.470. Both gains increase across these measured positions on each of the three checkpoints. This establishes the trend within that grid family, where prefix and suffix durations also change. The fixed-state final-interval experiment addresses propagation separately.

2.Horizon adjustment retains one-step generation

Horizon adjustment does not require giving up one-step sampling. We added one-step comparisons that retain one deployment velocity query, including examples on released checkpoints (Section 3.4; Appendix C).

The negative source result in the original four-step experiment concerns an update of duration δ=.25\delta=.25 followed by three further updates. A one-step sampler instead uses δ=1\delta=1 and has no downstream sampler. Although both begin at t=0t=0, they request different intervals and apply the prediction over different durations. The four-step source result therefore does not determine the one-step outcome.

On new 50k generation inputs, the one-step results are:

ModelScalar FIDCombination FIDHorizon
CIFAR, seven-checkpoint mean18.160217.00631.10
Released MeanFlow-XL/23.12723.0452.98
Released AlphaFlow2.36622.31921.02

Each row compares Scalar with the same scale plus Horizon on paired inputs. The CIFAR and public-model evaluations use different real reference sets, so the comparisons are within rows. CIFAR uses scale a=.98a=.98, while both public examples use a=1.03a=1.03. Public results each concern one released checkpoint.

The MeanFlow h=.98h=.98 query stays within the nominal time domain and also raises Inception Score (IS), from 217.755 to 221.751. CIFAR h=1.10h=1.10 and AlphaFlow h=1.02h=1.02 extrapolate beyond the nominal domain. For AlphaFlow, h=1.02h=1.02 was selected after scale calibration; the native-scale horizon sweep instead preferred h=1h=1. Its combination lowers IS from 242.765 to 241.309 despite the FID gain.

The one-step direction experiment provides a separate mechanism check: sample-matched corrections retain an advantage over strict direction permutations even after matching lengths, parallel components and angles. Thus both the practical comparison and the directional analysis can be carried out without introducing a midpoint. One velocity query means two backbone forwards for the CIFAR two-pass CFG implementation and one forward for the released models with built-in guidance.

Reviewer_TVnWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose a few-step appraoch called MeanFlow decoupling the prediction horizon from the update step, and they present the improvement on CIFAR-10

Critical strengths

The following protocol is the main strength, very frozen rules, prospective checkpoints, paired samples, multiple evaluators. The paper report their own negative results which is also important.

Critical weaknesses

Very weak scientific contribution and novelty. The paper is very long and the main text is hard to read because almost every claim is immediately qualified.

Questions for authors

What is the recommendation for a practitioner?

2 comments

1.Scientific contribution and novelty

Thank you for recognizing the paired evaluations, checkpoint replication and reporting of negative results. We agree that the earlier presentation did not explain the scientific contribution clearly enough.

The revised paper addresses three connected questions.

(1) Does horizon adjustment offer value beyond velocity scaling? On the stronger four-query CIFAR-10 grid, Combination reduces mean FID from 11.6727 for Scalar to 10.7930, with FID and KID improvements on all seven checkpoints. These are paired 10k evaluations at the same deployment query count. One-step comparisons on separate 50k inputs also show incremental gains, including on released MeanFlow and AlphaFlow checkpoints (Section 3.4; Appendix C). The practical result is that adding Horizon improves on Scalar at these calibrated operating points; Horizon alone does not outperform Scalar on the tested four-query CIFAR grids. Combination uses both scale and horizon calibration, so its total search cost is higher than either standalone search.

(2) Does this increment depend on the correction direction assigned to each sample? In one-step CIFAR generation at classifier-free guidance (CFG) 2, scale a=.98 and horizon h=1.10, we replace the correction's orthogonal direction using a same-class donor correction projected into the recipient's orthogonal space. The control preserves the recipient's correction length, signed parallel component and angle to the original velocity. Combination outperforms these strict controls by 3.230 FID on four checkpoints, with descriptive 95% checkpoint interval [1.295, 5.166]. Its gain over Scalar on the same inputs is 1.362. This identifies a sample-matching effect that these scalar geometric quantities do not capture (Section 4.1; Appendix D.2).

(3) Why is the few-step effect position dependent? For model-reference interventions, changing only the final interval [.75,1] reduces the midpoint-minus-quarter-point gain difference from 3.421 with one native update to 2.699 with two native half-steps and .073 with fourth-order Runge–Kutta (RK4) integration. All four confirmation checkpoints follow this ordering in FID and KID. The distribution analysis shows that the final sampler changes both the correction's feature response and its relation to the baseline distribution (Section 4; Appendix D.7).

These results move the contribution beyond a favorable midpoint setting: horizon changes have practical value at selected scales, their one-step benefit depends on sample-matched directions, and their few-step benefit depends strongly on the subsequent sampling map. The reference integrations use the model's instantaneous field and additional computation for diagnosis.

2.Length and readability

The nine-page main text now follows the scientific questions: practical value, mechanism and operating conditions.

Section 2 defines the control and its distinction from velocity scaling and grid changes. Section 3 establishes practical value at fixed deployment query counts, including the increment over calibrated scaling. Section 4 connects one-step direction matching to few-step propagation and endpoint distribution changes. Section 5 brings together operating conditions, competing temporal controls and the recommendation for practitioners.

Each main experiment now introduces its question, the quantities held fixed and the comparison being made. Detailed comparisons are grouped in thematic appendices: primary effects and transfer in Appendix B, calibration controls in C, mechanism experiments in D, evaluator robustness and qualitative examples in E, and predictive limitations and further comparisons in F. Appendix A includes a compact map of the experimental questions and controls.

We also condensed repeated qualifications and revised the abstract, introduction and conclusion to lead with the findings. Scope information now accompanies the relevant experiment: for example, the one-step h=1.10 query is identified as out of domain, and reference integrations are identified as additional-compute diagnostics.

3.Recommendation for practitioners

For practitioners, our recommendation is to calibrate the velocity scale and evaluate whether changing the prediction horizon improves that operating point on separate generation inputs. Section 5 gives this recommendation for a fixed deployment query budget.

Compare Local (the matched-horizon sampler), Horizon alone, Scalar alone and their Combination. The stronger four-query CIFAR grid illustrates why: Scalar is a stronger standalone control than Horizon, but adding Horizon to Scalar reduces mean FID from 11.6727 to 10.7930. The choice should also reflect the application's metrics. For example, the one-step AlphaFlow combination improves FID over Scalar but lowers Inception Score (IS), so we report both.

Horizon does not require extra deployment queries or a midpoint. Released MeanFlow supports an in-domain one-step example with h=.98; the CIFAR one-step h=1.10 example instead queries beyond the nominal time domain. Calibration has a separate cost: Combination includes both scale and horizon searches. Query-start and grid alternatives are also useful comparisons when calibration resources permit.

Fixing query count is the experimental condition used to isolate horizon effects, not a restriction on applications. If compute is flexible, query count, grid, scale and horizon can be calibrated together. We identify that broader optimization as future work in an additional response; the current recommendation is supported by comparisons at the tested operating points.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Additional response: Joint calibration across compute budgets

Our fixed-query comparisons isolate the effect of prediction horizon from improvements due to additional model evaluations. Fixing query count serves this experimental purpose; applications can vary the available computation.

As future work, we plan to study joint calibration of query count, sampling grid, velocity scale and prediction horizon. We will evaluate the resulting quality–compute tradeoffs with both deployment latency and calibration cost reported, using separate inputs for parameter selection and evaluation. This would test whether the incremental value of horizon adjustment persists when the other sampling choices are optimized together.

The existing results establish gains at the tested operating points. They include comparisons across selected query budgets, but not a joint optimization of all four choices.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.