acceptodds
Under review as a conference paper at ICLR 2027

Sharpen Your Flow: Sharpness-Aware Sampling for Flow Matching

Abstract

Flow matching models generate samples by numerically integrating a learned velocity field, with each integration step requiring a neural network evaluation. Fast generation therefore requires using a small fixed evaluation budget effectively: the key question is not only how to integrate the flow, but where the sampler should spend its steps. We propose SharpEuler, a training-free sampler that profiles a pretrained model offline by estimating where the learned velocity field changes most rapidly along calibration trajectories. This finite-difference estimate defines a solver-aware sharpness profile, which is smoothed and converted by a quantile transform into a timestep grid for any desired inference budget. At test time, sampling remains ordinary Euler integration with the same number of model evaluations as a uniform schedule. We justify SharpEuler using three principles: a numerical principle identifying trajectory acceleration as the leading source of Euler discretization error, a variational principle deriving sharpness-based power-law timestep densities, and a statistical guarantee showing that the finite-sample calibrated sampler is stable at the terminal distribution level. Our experiments show that SharpEuler improves sample quality at fixed budgets, reducing inter-mode leakage and increasing mode coverage.

Primary area:
generative models
Keywords:
flow matching, few-step sampling

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_t3G7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SharpEuler, a training-free timestep scheduler for pretrained flow matching models. The method estimates trajectory acceleration from offline velocity differences and converts the smoothed profile into a fixed Euler grid. The analysis motivates a square-root allocation, ρ(t)∝E∣x¨(t)∣\rho(t)\propto\sqrt{\mathbb{E}|\ddot{x}(t)|}, through an Euler local-error objective. Experiments on synthetic distributions, FLUX.1-dev, and Pyramid Flow report improved low-budget sampling and closer agreement with high-step references.

Critical strengths
  • The method is simple and practically attractive. A single offline calibration profile supports different sampling budgets without retraining or additional inference evaluations.
  • The Euler-error analysis provides a clear motivation for acceleration-based allocation. The FLUX results show substantial improvements in reference agreement, supported by the additional 1,000-prompt evaluation.
Critical weaknesses
  • The comparison with closely related scheduling methods remains limited. FLUX is evaluated only against the default and static shifted schedules, while OSS-ave appears only on synthetic datasets. Since OSS [1] already studies calibrated schedules for FLUX, a comparison at matched NFE, together with calibration costs, is important to establish the practical advantage.
  • The statistical guarantee does not directly cover the full implementation. Theorem 2 assumes an idealized interval allocation. Errors from finite differences and inverse-CDF interpolation are not explicitly bounded, especially when inference uses far fewer steps than calibration.
  • The FLUX results use an additional final-step correction without a separate ablation. Its contribution to the gains is therefore unclear. Calibration also uses a shifted grid, while the algorithms assume uniform spacing. The handling of interval widths and endpoints needs clarification.

[1] Optimal Stepsize for Diffusion Sampling, Pei et al., arXiv 2025

Questions for authors
  • How does Theorem 2 account for finite-difference and inverse-CDF discretization errors when the inference budget is much smaller than the calibration grid?
  • For FLUX, how are nonuniform calibration intervals and the final-step replacement handled? Are results available without the tail-alignment correction?
3 comments

We thank the reviewer for the careful review and constructive comments.

Comparison with OSS-ave and KTS on FLUX (W1). We have extended the FLUX experiments to include OSS-ave and KTS at matched inference budgets, with offline calibration costs reported separately. We evaluate all methods on the same 1,000 COCO captions, using identical evaluation prompts, generation seeds, and 50-step reference images. All experiments use FLUX.1-dev at 1024×10241024\times1024 resolution, with inference budgets of B=8B=8 and B=12B=12 model evaluations. For methods requiring offline calibration, we use 64 calibration prompts and 50 profiling or teacher steps. KTS uses the balanced hyperparameter setting reported in its paper.

MethodB=8 RMSE ↓CLIP ↑FID ↓B=12 RMSE ↓CLIP ↑FID ↓
Default FLUX53.690.30842.8949.630.30737.02
OSS-ave38.730.30837.7333.650.30725.92
KTS53.830.30843.2149.790.30837.06
SharpEuler32.580.30733.0733.010.30725.31

Reference FID is computed against the fixed set of 1,000 images generated using the 50-step FLUX pipeline.

SharpEuler achieves the lowest RMSE and reference FID at both budgets. Compared with OSS-ave, at B=8B=8, RMSE decreases from 38.73 to 32.58 and reference FID from 37.73 to 33.07. At B=12B=12, RMSE decreases from 33.65 to 33.01 and reference FID from 25.92 to 25.31. CLIP scores remain close across methods.

Offline calibration cost. SharpEuler estimates a single continuous sharpness profile that can be reused across inference budgets, whereas our OSS-ave implementation performs a separate schedule search for each budget.

MethodCalibration promptsProfiling / teacher stepsOffline model evaluations
SharpEuler64503,200 total, reused for both budgets
OSS-ave, B=8B=8645028,800
OSS-ave, B=12B=12645041,600
KTS——0

All calibration and search computations were performed on NVIDIA A100 GPUs. SharpEuler requires 9$\times$ fewer offline model evaluations than OSS-ave at B=8B=8 and 13$\times$ fewer at B=12B=12. Furthermore, the same SharpEuler profile supports both inference budgets, whereas OSS-ave requires a separate search. These offline evaluations are not included in the matched inference NFE. We will include these comparisons and implementation details in the appendix.

Statistical guarantee and practical discretization (W2, Q1). We agree that Theorem 2 abstracts from some implementation details. It explicitly absorbs finite-difference and calibration-grid biases into the residual term C2C_2, while the inverse-CDF construction is the deterministic realization of the calibrated density: each interval carries mass 1/B1/B, so the practical grid implements the same allocation principle analyzed theoretically. When B≪NB\ll N, the calibration profile is still estimated on the fine grid NN, while the smaller BB simply makes the resulting Euler approximation coarser, as reflected by the O(1/B)O(1/B) term; the theorem is therefore a stability guarantee rather than a tight few-step error bound. Note that a theorem covering every implementation detail would necessarily depend on model- and task-specific choices such as time parameterization, interpolation, endpoint handling, and multistage sampling, so we intentionally isolate the model-independent calibration error and treat these deterministic approximations separately. We will clarify this scope in the revision.

Tail-alignment ablation (W3, Q2). We conducted a separate ablation to isolate the contribution of the FLUX tail-alignment correction. We use the same 1,000 COCO evaluation captions, original calibration profile, generation seeds, and inference budgets, changing only whether the final step is aligned with the native FLUX pipeline.

MethodB=8 RMSE ↓\downarrowCLIP ↑\uparrowFID ↓\downarrowB=12 RMSE ↓\downarrowCLIP ↑\uparrowFID ↓\downarrow
Native FLUX53.690.30842.8949.630.30737.02
SharpEuler (with tail alignment)32.580.30733.0733.010.30725.31
SharpEuler (without tail alignment)33.820.30733.6833.070.30725.36

Reference FID is computed against the fixed set of 1,000 images generated with the 50-step FLUX pipeline.

SharpEuler retains substantial improvements over native FLUX even without tail alignment. At B=8B=8, removing the correction changes the reference FID only from 33.07 to 33.68, compared with 42.89 for native FLUX. At B=12B=12, reference FID changes from 25.31 to 25.36, compared with 37.02 for native FLUX.

These results show that the principal gains arise from the calibrated timestep allocation rather than the final-step correction. Tail alignment provides a modest additional improvement and was introduced primarily to preserve perceived sharpness and fine visual details near the low-noise endpoint as noted in Appendix D.4. We will include this ablation and paired visual examples in the appendix.

Nonuniform calibration intervals and endpoint handling (W3, Q2). SharpEuler's construction in Section 2 is defined for a general descending calibration grid,

1=s0>s1>⋯>sN=0,1=s_0>s_1>\cdots>s_N=0,

rather than requiring uniformly spaced timesteps. In particular, Eq. (4) divides consecutive velocity differences by the local interval width, defined as the average of the two adjacent calibration intervals. The estimate therefore accounts for nonuniform spacing rather than assuming a constant timestep. For FLUX, calibration uses the default 50-step pipeline described in Appendix D.4, retaining its shifted time coordinates. The sharpness estimates are assigned to their corresponding time midpoints before constructing the interpolated cumulative profile.

As specified in Appendix D.4, tail alignment replaces the final SharpEuler step with the corresponding native FLUX pipeline step, preserving the pipeline's low-noise tail, where the calibrated finite-difference profile is empirically noisier. The initial and terminal endpoints remain fixed, and the correction introduces no additional model evaluations. The sampler therefore continues to use exactly BB network evaluations. We will make these implementation details and the new experimental ablations explicit in the revised manuscript.

We would like to thank the reviewer again for their time and constructive feedback. We believe our responses and additional experiments address the main points raised in the review. We would be happy to clarify any remaining questions during the discussion period.

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

This paper proposes SharpEuler, a training free sampler that can identify the areas with rapid velocity changes along calibration trajectories, and defines a slover aware sharpness profile that can be converted into a timestep grid for a desired inference step. At test time, sampler remains ordinary Euler integration as a uniform schedule. Experiment are conducted on synthesis and real world data to validate the effectiveness of the proposed approach.

Critical strengths

The proposed approch is support by three theoretical principles.

The experiments are sound to support the proposed approach.

Critical weaknesses

The experiments should provide some other baselines to show the advantage of the proposed approach, for example higher order sampler.

Questions for authors

No

1 comment

We thank the reviewer for the positive assessment of our work and the suggestion to include higher-order samplers.

Appendix E.4 evaluates sharpness-based scheduling with Heun’s method, comparing calibrated and uniform grids at matched NFE. To further strengthen the empirical evaluation, we have extended the FLUX experiments to include OSS-ave and KTS at matched inference budgets, with offline calibration costs reported separately. We have also added quantitative video-quality and temporal-consistency evaluations. These results will be included in the revised manuscript.

Reviewer_2QpvOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SharpEuler, a training-free timestep scheduling method for pretrained flow-matching models. It estimates trajectory acceleration from finite differences of model velocities on offline calibration trajectories, smooths the resulting profile, and uses a power-law weighting and quantile transform to construct a shared timestep grid for a chosen evaluation budget. Inference retains ordinary Euler updates and the original model weights. The paper develops numerical, variational, and statistical analyses relating trajectory acceleration to Euler error, timestep allocation, and finite-sample stability.

Validation covers synthetic distributions, FLUX.1-dev text-to-image generation, and qualitative PyramidFlow text-to-video examples. Synthetic experiments examine distributional fidelity and coverage, while FLUX experiments measure agreement with 50-step references and prompt alignment, with an additional 1,000-prompt distributional evaluation. The reported results show improved low-budget sampling on the synthetic tasks and closer agreement with high-step image references at matched model-evaluation budgets, while maintaining similar CLIP scores.

Critical strengths

1. Simple deployment with fixed inference cost. SharpEuler preserves the pretrained model and ordinary Euler updates, adding neither auxiliary networks nor online error estimation. A single offline profile can produce schedules for different sampling budgets. This makes the approach practical for large models, where each additional network evaluation is costly.

2. Interpretable scheduling rationale with supporting empirical evidence. The acceleration-based signal explicitly connects timestep allocation to Euler's leading local discretization error, and the variational analysis motivates a direct density-based schedule construction. The synthetic and FLUX experiments report improvements over the evaluated baseline schedules at matched NFE, including closer agreement with high-step image references while maintaining similar prompt alignment. These results provide initial evidence of the scheduling rule's practical usefulness.

Critical weaknesses

1. Calibration-set construction is heuristic and insufficiently specified. The paper describes broad semantic coverage but gives no reproducible selection rule for the 64 calibration prompts or sensitivity analysis of calibration-set choice on FLUX. The source and sampling protocol for the 1,000 held-out prompts are also unclear, leaving the dependence on calibration-set choice and applicability to new prompt distributions insufficiently established. Although FLUX and PyramidFlow use 64 and eight calibration prompts respectively, the criteria for choosing these sets and sizes, or reusing prompts across models, are not explained.

2. Evidence against stronger scheduling baselines is insufficient on text-to-image and text-to-video models. FLUX reports RMSE, FID/KID, and CLIP, but compares only with the default pipeline and a static shift; comparisons with OSS-ave and KTS are confined to synthetic data. PyramidFlow provides only qualitative comparisons with uniform Euler, without quantitative video-quality or temporal-consistency results. Comparisons with applicable recent training-free or learned scheduling methods, at matched NFE and with calibration/training costs reported, would strengthen the evidence: even a modest quality gain, or comparable quality at lower offline cost, would make the practical value more convincing.

Questions for authors
  1. Appendix E.6 states that the 1,000 evaluation prompts are disjoint from the 64 calibration prompts. Please clarify the sources and reproducible selection procedures for the calibration set and both evaluation sets, including the relationship between the 80- and 1,000-prompt evaluations. What motivates using 64 calibration prompts for FLUX, and how sensitive are the results to calibration-set choice? When transferring to another model or a public prompt source such as COCO, can the original calibration prompts be reused with separate calibration, or must a new set be selected? Please specify the selection and size criteria for any new calibration set, and whether the existing evaluation demonstrates transfer to a different prompt distribution.

  2. Could you compare with applicable recent training-free or learned scheduling methods (e.g., HSO [1] for training-free schedule optimization, LD3 [2] for learned offline discretization, or ETS [3] beyond the existing synthetic-data comparison, subject to model and solver compatibility) on text-to-image models at matched NFE, reporting generation quality and offline calibration/training costs? These comparisons would help establish whether the method offers quality gains or comparable quality at lower offline cost.

References

[1] Zhu A, Su R, Zhao Q, et al. Hierarchical schedule optimization for fast and robust diffusion model sampling[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2026, 40(16): 13907-13915.

[2] Tong V, Hoang T D, Liu A, et al. Learning to discretize denoising diffusion odes[C]//International Conference on Learning Representations. 2025, 2025: 47244-47282.

[3] Stancevic D, Handke F, Ambrogioni L. Entropic time schedulers for generative diffusion models[J]. Advances in Neural Information Processing Systems, 2026, 38: 44222-44252.

4 comments

We thank the reviewer for the careful review and constructive comments.

Calibration-set construction and transferability (W1, Q1). We have now clarified the construction of the calibration and evaluation prompt sets. We will release the exact prompt lists, sampling scripts, and experimental settings needed to reproduce our results.

The original 64 FLUX calibration prompts were selected from an LLM-generated prompt pool spanning different semantic categories, including landscapes, urban scenes, wildlife, and abstract content. The 80-prompt and 1,000-prompt evaluation sets were also drawn from this pool, with the calibration prompts excluded from both evaluations. The 80-prompt evaluation was used for paired RMSE, CLIP, and reference FID comparisons, while the 1,000-prompt evaluation was used for the larger-scale KID and FID-1k measurements. Both sets exclude all 64 calibration prompts. We will document the exact prompt identifiers and the relationship between these evaluation sets in the appendix.

Following your suggestion, we additionally conducted experiments using COCO2014 validation captions to provide a reproducible public-data evaluation. We randomly selected calibration captions and fixed a set of 1,000 evaluation captions, ensuring that calibration and evaluation were disjoint at the image level. We also constructed five mutually image-disjoint random sets of 64 calibration captions, all excluding the evaluation images. We will release the corresponding image and caption IDs and the sampling scripts.

We evaluated three calibration profiles on the same fixed COCO evaluation set: (i) the profile calibrated using our original LLM-generated prompts, transferred unchanged to COCO; (ii) a profile calibrated on a random set of 64 COCO captions; and (iii) a profile calibrated on a second, disjoint set of 64 COCO captions. All experiments use FLUX.1-dev at 1024×1024 resolution and the same evaluation prompts and noise seeds.

CalibrationB=8 RMSE ↓CLIP ↑FID ↓B=12 RMSE ↓CLIP ↑FID ↓
Native FLUX53.690.30842.8949.630.30737.02
Original profile32.580.30733.0733.010.30725.31
COCO-calibrated, set 133.730.30733.0729.650.30625.31
COCO-calibrated, set 233.960.30834.3532.820.30726.07

Reference FID is computed against the fixed set of 50-step FLUX-generated reference images.

The original profile, calibrated on LLM-generated prompts, retains substantial improvements when applied unchanged to COCO. At B=8, it reduces RMSE from 53.69 to 32.58 and reference FID from 42.89 to 33.07. Recalibrating on either of the two random COCO subsets also preserves these gains. This provides evidence that the improvement is not restricted to the original calibration prompt distribution.

We also studied the stability of the profiles across the five independently selected COCO calibration sets. For the normalized scheduling-weight profiles, the average pointwise standard deviation is 0.000233, and the maximum pointwise standard deviation is 0.001350. The resulting profiles closely overlap, indicating limited variation across these prompt draws.

Choice of calibration-set size. We chose 64 prompts as a conservative practical setting, rather than a theoretically required minimum or an optimized value. We additionally evaluated FLUX using 16, 32, and 64 calibration prompts. Downstream performance remains similar across these sizes: reference FID ranges from 33.45 to 33.59 at B=8 and from 25.07 to 25.36 at B=12, compared with 42.89 and 37.02 for the native pipeline. Across 50 random subsets of the pooled calibration trajectories at each size, the average pointwise standard deviation of the normalized scheduling weights decreases from 0.000605 with 16 prompts to 0.000284 with 64 prompts. We will include this ablation and the corresponding profile plots in the appendix.

SharpEuler's calibrated profile depends on the model's learned velocity field. Therefore, transferring the method to another model requires re-estimating its profile, although the same text prompts can in principle be reused if they are suitable for that model. For Pyramid Flow, we separately calibrate a three-stage profile using eight prompts and 60-step trajectories per stage, as described in Appendix D.6. We have not established a theoretically optimal calibration-set size across models. The present experiments support the practical adequacy of 64 prompts for FLUX, while calibration-size selection for other models remains empirical.

Comparison with stronger scheduling baselines on FLUX (W2, Q2). We have extended the FLUX experiments to OSS-ave and KTS at matched inference budgets, reporting offline calibration costs separately.

Regarding the additional methods suggested by the reviewer, their published implementations target different sampling formulations. HSO is implemented for Stable Diffusion 2.1 and requires model-specific adaptation of its noise schedule. LD3 learns timestep discretizations through offline optimization; although its authors report promising preliminary FLUX results, the corresponding learned FLUX timesteps have not been publicly released. ETS derives its schedule from entropy-based quantities for diffusion models, and its released implementation targets EDM/EDM2. Applying these methods to FLUX.1-dev therefore requires additional implementation and validation to ensure a meaningful comparison.

We prioritize OSS-ave, which directly optimizes timestep schedules and has already been studied on FLUX, alongside KTS as a complementary training-free acceleration method. Our original comparisons were limited to synthetic distributions; we now extend them to FLUX and report offline costs separately.

We evaluate all methods on the same 1,000 COCO captions, using identical evaluation prompts, generation seeds, and 50-step reference images. All experiments use FLUX.1-dev at 1024×1024 resolution, with matched inference budgets of B=8 and B=12 model evaluations. For methods requiring offline calibration, we use 64 calibration prompts and 50 profiling or teacher steps. KTS uses the balanced hyperparameter setting reported in its paper.

MethodB=8 RMSE ↓CLIP ↑FID ↓B=12 RMSE ↓CLIP ↑FID ↓
Default FLUX53.690.30842.8949.630.30737.02
OSS-ave38.730.30837.7333.650.30725.92
KTS53.830.30843.2149.790.30837.06
SharpEuler32.580.30733.0733.010.30725.31

SharpEuler achieves the lowest RMSE and reference FID at both budgets, while CLIP scores remain comparable across methods. OSS-ave optimizes the timestep schedule, whereas KTS modifies the velocity field during sampling. We include both to broaden the comparison with training-free acceleration methods.

Offline calibration cost. SharpEuler estimates one continuous sharpness profile that can be reused across inference budgets. In contrast, our OSS-ave implementation performs a separate schedule search for each budget. The offline costs are:

MethodCalibration promptsProfiling / teacher stepsOffline model evaluations
SharpEuler64503,200 total, reused for both budgets
OSS-ave, B=8645028,800
OSS-ave, B=12645041,600
KTS——0

All calibration and search computations were performed on NVIDIA A100 GPUs. SharpEuler uses 9× fewer offline model evaluations than our OSS-ave search at B=8 and 13× fewer at B=12. Moreover, the same SharpEuler profile supports both inference budgets, whereas OSS-ave requires a separate search. These offline evaluations are excluded from the matched inference NFE. We will include the results and implementation details in the appendix. These experiments broaden the realistic-model comparisons.

Quantitative video evaluation (W2). We also agree that the original Pyramid Flow results were qualitative. We therefore supplemented them with three reference-free metrics from the official VBench implementation [1], which are also used in the Pyramid Flow literature:

  • Video Quality (MUSIQ): Measures technical image quality, including blur, noise, and exposure distortions.
  • Aesthetic Quality (LAION): Measures visual appeal, including composition, color, and overall aesthetics.
  • Subject Consistency (DINO): Measures consistency of subject appearance across frames.

We report the mean scores across the paired baseline–SharpEuler video comparisons at the same generation budget (B = 20).

MethodVideo Quality (MUSIQ) ↑Aesthetic Quality (LAION) ↑Subject Consistency (DINO) ↑
Default Pyramid Flow0.47030.58960.9779
SharpEuler0.62250.65620.9901

SharpEuler improves all three metrics, supporting the qualitative observations in Figure 4. Scores use VBench's output scales. We will include the per-example results and evaluation configuration in the appendix.

References

[1] Huang et al. VBench: Comprehensive Benchmark Suite for Video Generative Models. CVPR 2024. https://arxiv.org/abs/2311.17982

[2] Jin et al. Pyramidal Flow Matching for Efficient Video Generative Modeling. ICLR 2025. https://arxiv.org/abs/2410.05954

Thank you for the detailed response. The additional experiments address most of my concerns. While the choice of calibration prompts remains empirical, the COCO results show that SharpEuler works well with randomly sampled calibration sets and that the original FLUX schedule transfers to a different prompt source. At matched inference budgets, SharpEuler improves RMSE and reference FID over OSS-ave while maintaining comparable CLIP scores and requiring substantially fewer offline model evaluations.

I find the method interesting and am raising my score to 3. I would still welcome experiments on additional text-to-image or text-to-video flow-matching models to better establish how broadly the method applies.

We thank the reviewer for taking the time to read our responses and additional experiments. We appreciate the updated score. If accepted, we will extend our evaluation to additional generative models in the camera-ready version.

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

SharpEuler uses offline trajectory measurements to allocate more sampling steps where a pretrained flow model's velocity changes rapidly. The resulting schedule is reused at inference with ordinary Euler integration and a fixed evaluation budget. The paper motivates the design through numerical error analysis and evaluates it on synthetic distributions, FLUX image generation, and a video example.

Critical strengths
  1. The method is simple to deploy: it changes the timestep schedule without retraining the model or adding inference evaluations. The connection between trajectory acceleration and Euler error gives the design a clear rationale.
  2. The FLUX results show better agreement with the high-step reference at matched budgets, including an evaluation on 1,000 held-out prompts. The synthetic studies also examine calibration size and the trade-off between fidelity and coverage.
Critical weaknesses
  1. The reliability of a single calibration profile on large models is not fully characterized. Held-out prompt results are encouraging, but the calibration-size study is limited to synthetic data. It remains unclear how much FLUX performance varies with the calibration prompts and noise seeds.
  2. Comparisons with alternative timestep schedules, such as OSS-ave and KTS, are confined to synthetic experiments. The FLUX comparisons use the default and a shifted schedule, while the video evidence is qualitative. This leaves the advantage over competing scheduling methods on realistic generation tasks less established.
Questions for authors
  1. How much do the FLUX results change when the calibration prompts and seeds are resampled? Are there prompt groups for which the shared schedule consistently performs worse?
  2. How does SharpEuler compare with OSS-ave or KTS on FLUX at matched inference budgets, with offline calibration costs reported separately?
3 comments

We thank the reviewer for the careful review and constructive comments.

Robustness of FLUX calibration (W1, Q1). We agree that the reliability of a shared calibration profile on large models deserves further investigation. To provide additional empirical evidence, we extended the calibration-size study to FLUX.1-dev, studying both the stability of the resulting scheduling profiles and their effect on downstream generation quality. Following another reviewer's suggestion to use a public prompt source, we randomly selected COCO captions for calibration and used a fixed set of 1,000 evaluation captions, disjoint from calibration at the image level. We varied the calibration size from 16 to 64 prompts and evaluated the resulting schedules at two inference budgets.

Calibration promptsB=8 RMSE ↓CLIP ↑FID ↓B=12 RMSE ↓CLIP ↑FID ↓
Default FLUX53.690.30842.8949.630.30737.02
1633.240.30833.4533.080.30725.36
3233.450.30833.5932.980.30725.07
6433.440.30833.4933.000.30725.10

Reference FID is computed against the fixed set of 50-step FLUX-generated reference images.

All three calibration sizes improve RMSE and reference FID over the default pipeline at both budgets, with little variation between 16, 32, and 64 prompts. For example, at B=12, reference FID ranges only from 25.07 to 25.36, compared with 37.02 for the default pipeline. The scheduling profiles are also stable. Across 50 random subsets of the pooled calibration trajectories at each size, the average pointwise standard deviation of the normalized scheduling weights decreases from 0.000605 with 16 prompts to 0.000284 with 64 prompts.

To further study sensitivity to the particular calibration prompts, we additionally constructed five mutually image-disjoint random sets of 64 COCO captions, all disjoint from the evaluation images. Across the five resulting normalized scheduling-weight profiles, the average pointwise standard deviation is 0.000233, and the maximum pointwise standard deviation is 0.001350. The profiles closely overlap, indicating limited variation in the calibrated schedule across these prompt draws.

We will add these results and the corresponding profile plots to the appendix. These experiments provide evidence of stability across calibration-prompt selections and sizes. We also evaluate profiles recalibrated on two disjoint sets of 64 COCO captions, alongside the original profile used in the paper, calibrated on 64 LLM-generated prompts and transferred unchanged to COCO evaluation set.

CalibrationB=8 RMSE ↓CLIP ↑FID ↓B=12 RMSE ↓CLIP ↑FID ↓
Native FLUX53.690.30842.8949.630.30737.02
Original profile32.580.30733.0733.010.30725.31
COCO-calibrated, set 133.730.30733.0729.650.30625.31
COCO-calibrated, set 233.960.30834.3532.820.30726.07

All three retain substantial improvements over native FLUX and achieve broadly similar downstream scores. Together with the closely overlapping calibration-profile plots, these downstream results provide empirical evidence of robustness to calibration-set choice. A full characterization of sensitivity to calibration prompts and noise seeds, including systematic differences across prompt groups, remains an open question for future work.

Comparison with OSS-ave and KTS on FLUX (W2, Q2). We agree that comparisons with competing methods on a realistic generation task are important. Thus, we extended the FLUX experiments to OSS-ave and KTS at matched inference budgets, reporting offline calibration costs separately.

We evaluate all methods on the same 1,000 COCO captions using identical evaluation prompts, generation seeds, and 50-step reference images. All experiments use FLUX.1-dev at 1024×1024 resolution, with inference budgets of B=8 and B=12. For methods requiring offline calibration, we use 64 calibration prompts and 50 profiling or teacher steps. KTS uses the balanced hyperparameter setting reported in its paper and requires no offline calibration.

MethodB=8 RMSE ↓CLIP ↑FID ↓B=12 RMSE ↓CLIP ↑FID ↓
Default FLUX53.690.30842.8949.630.30737.02
OSS-ave38.730.30837.7333.650.30725.92
KTS53.830.30843.2149.790.30837.06
SharpEuler32.580.30733.0733.010.30725.31

SharpEuler achieves the lowest RMSE and FID at both budgets. At B=8, it reduces RMSE from 38.73 to 32.58 and FID from 37.73 to 33.07 compared with OSS-ave. At B=12, the improvements are smaller but consistent: RMSE decreases from 33.65 to 33.01, and reference FID from 25.92 to 25.31. CLIP scores remain close across methods. These comparisons use matched inference NFE. OSS-ave optimizes the timestep schedule, as SharpEuler does, whereas KTS modifies the velocity field during sampling. We include KTS as a complementary training-free acceleration baseline.

Offline calibration cost. SharpEuler estimates a single continuous sharpness profile that can be reused across inference budgets, whereas our OSS-ave implementation performs a separate schedule search for each budget. The offline costs are:

MethodCalibration promptsProfiling / teacher stepsOffline model evaluations
SharpEuler64503,200 total, reused for both budgets
OSS-ave, B=8645028,800
OSS-ave, B=12645041,600
KTS——0

All calibration and search computations were performed on NVIDIA A100 GPUs. SharpEuler uses 9× fewer offline model evaluations than OSS-ave search at B=8 and 13× fewer at B=12. Moreover, the same SharpEuler profile supports both budgets, whereas OSS-ave requires a separate search. We will add these comparisons and implementation details to the appendix.

Quantitative video evaluation (W2). We also agree that the original Pyramid Flow experiment provided only qualitative evidence. To supplement it, we evaluated the generated videos using the official VBench implementation [1], including three reference-free metrics:

  • Video quality (MUSIQ): Measures technical image quality, including blur, noise, and exposure distortions.
  • Aesthetic quality (LAION): Measures visual appeal, including composition, color, and overall aesthetics.
  • Subject consistency (DINO): Measures the consistency of visual subject features across video frames.

We report the mean scores across baseline–SharpEuler video comparisons at the same generation budget (B = 20).

MethodVideo quality (MUSIQ) ↑Aesthetic quality (LAION) ↑Subject consistency (DINO) ↑
Default Pyramid Flow0.47030.58960.9779
SharpEuler0.62250.65620.9901

SharpEuler improves all three metrics in these comparisons, providing quantitative support for the qualitative observations in Figure 4. These scores are reported on the scales returned by VBench. We will include the per-example results and evaluation configuration in the appendix.

References

[1] Huang et al. VBench: Comprehensive Benchmark Suite for Video Generative Models. CVPR 2024. https://arxiv.org/abs/2311.17982

[2] Jin et al. Pyramidal Flow Matching for Efficient Video Generative Modeling. ICLR 2025. https://arxiv.org/abs/2410.05954

We would like to thank the reviewer again for their time and constructive feedback. We believe our responses and additional experiments address the main points raised in the review. We would be happy to clarify any remaining questions during the discussion period.

Comments on the paper

1 comment

We thank the reviewers for their constructive feedback. We have uploaded a revised manuscript with additional experiments, ablations, and implementation details. Below we summarize the main updates:

  • Expanded FLUX evaluation: We added experiments on 1,000 held-out COCO captions, comparing SharpEuler with OSS-ave, KTS, uniform Euler, and shifted schedules at matched inference budgets. SharpEuler achieves the lowest RMSE and reference FID at both 8 and 12 NFE, while requiring 9× and 13× fewer offline model evaluations than OSS-ave, respectively.

  • Calibration robustness and transferability: We extended the FLUX calibration-size study to 16, 32, and 64 COCO captions, analyzed profile variability across independent calibration sets, and evaluated the transfer of the original LLM-prompt-calibrated schedule to COCO. The results show stable scheduling profiles and consistent improvements across calibration settings.

  • Additional ablations and clarifications: We added a tail-alignment ablation showing that SharpEuler retains substantial improvements without the final-step correction. We also clarified the handling of nonuniform calibration grids, endpoint alignment, and the relationship between the statistical guarantee and the practical implementation.

  • Quantitative video evaluation: We supplemented the qualitative Pyramid Flow results with VBench evaluations of imaging quality (MUSIQ), aesthetic quality (LAION), and subject consistency (DINO). SharpEuler improves all three metrics at the same inference budget.

We hope these additions address the reviewers' questions and would be happy to provide further clarification during the remaining discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.