acceptodds
Under review as a conference paper at ICLR 2027

VeloForesight: Structured Admission for Training-Free Flow Acceleration

Abstract

Diffusion transformers generate high-quality videos, but repeated network evaluations make inference expensive. Predicting velocities from sparse exact evaluations can reduce this cost. However, forecasts based on sparse evaluations can become unreliable over longer prediction horizons, and their errors affect subsequent sampling states. In this paper, we introduce VeloForesight, a training-free sampler built around a structured predictor of velocity changes. The predictor estimates trend and curvature from past exact evaluations, controls their contributions separately, and progressively attenuates curvature over longer forecast horizons. Relative-norm bounds limit the size of the predicted changes before they enter the native scheduler. Across image and video generation, VeloForesight achieves up to speedup on Wan2.1-14B and on SD3.5 Large relative to 50-step sampling.

Primary area:
generative models
Keywords:
diffusion models, video generation, training-free acceleration, velocity forecasting, flow matching, diffusion sampling

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_4DEROpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VeloForesight, a training-free sampling acceleration method for flow-based image and video generation models. The method reduces the number of network evaluations by extrapolating velocity estimates from previously computed outputs. It leverages first- and second-order temporal differences between recent exact velocity evaluations to estimate the velocity at skipped sampling steps. To prevent potentially unstable extrapolation, the method introduces a reliability control mechanism.

The proposed framework further incorporates residual calibration and clean-state feedback to correct the extrapolated velocities, as well as a noise-progress-based strategy for allocating exact model evaluations. The authors evaluate the method on image and video generation models, comparing it against existing acceleration methods, including TaylorSeer, Spectrum, and an adapted version of VDE. The experimental results demonstrate improved inference efficiency and generally competitive or better fidelity compared with the evaluated baselines. Ablation studies further investigate the effectiveness of the proposed reliability control and auxiliary feedback mechanisms.

Critical strengths
  • The proposed approach is simple and easy to follow. Extrapolating velocities using first- and second-order temporal differences, followed by reliability control and calibration, appears to be a reasonable approach to reducing expensive network evaluations. The method also requires no additional training and preserves the original denoiser and sampling scheduler.
  • The experiments cover multiple image and video generation backbones and demonstrate improvements over the considered baselines.
Critical weaknesses
  • The idea of leveraging historical model outputs to extrapolate future values and avoid expensive network forward passes has already been explored in prior work, such as TaylorSeer and Spectrum. In particular, TaylorSeer employs Taylor-series-based extrapolation using finite differences of previously computed features, which is conceptually very similar to the proposed velocity predictor. The main distinction appears to be the prediction target: VeloForesight extrapolates the final velocity outputs rather than intermediate network features. Moreover, FastFlow has already explored direct velocity extrapolation using finite differences. Although the proposed structured reliability control introduces additional mechanisms for regulating the extrapolated updates, the paper does not sufficiently establish how these mechanisms constitute a substantial technical advancement over existing reliability estimation and correction mechanisms.
  • The correction mechanisms in Sec. 3.3 appear to be heuristic and sensitive to the choice of hyperparameters.
  • Despite comparing against VDE (CVPR 2026), the main experiments do not include comparisons with FastFlow (ICLR 2026), which appears particularly relevant given its use of finite-difference-based velocity extrapolation. While Appendix G provides a same-state diagnostic comparison using an equation-level adaptation of FastFlow, this is not equivalent to a comparison with the complete FastFlow. Given the substantial methodological overlap, a comparison against it would help establish the effectiveness of the proposed method.
  • One of the main contributions of this work is the introduction of reliability control and auxiliary feedback mechanisms for correcting extrapolated velocities. However, the experiments do not include comparisons with the existing methods discussed in the "Prediction reliability and correction" paragraph of the Related Work section.
  • Several expressions could be clarified to improve readability. For example:
    • L91: "We introduce a multi-anchor velocity predictor whose complete scheduler-facing change proposal is controlled before the sampler update..." The term "scheduler-facing change proposal" is somewhat artificial and could be replaced with a more direct description of the bounded velocity update supplied to the scheduler.
    • L102: The term "local comparisons" is ambiguous. It would be helpful to clarify whether this refers to locally reproduced baseline implementations, matched evaluation environments, or both.
    • L235: "At an exact step a, we reconstruct ... from the old anchor history, before inserting the new velocity." The phrase "inserting the new velocity" could be made more explicit. Presumably, this refers to appending the newly evaluated velocity to the exact-anchor history, rather than modifying the sampling state.
Questions for authors
  • Could the authors clarify the fundamental technical differences between the proposed velocity extrapolation method and prior approaches such as TaylorSeer, Spectrum, and FastFlow? What is the main methodological novelty of the proposed predictor beyond applying separate gains, horizon-dependent attenuation, and norm bounds?
  • Could the authors explain why comparisons with FastFlow are not included in the main experiments? Given the similarity between the two approaches, a comparison under matched evaluation budgets and computational settings would be particularly informative.
  • Could the authors provide additional empirical comparisons with the prediction reliability and correction methods discussed in the Related Work section, such as Confidence-Gated Taylor, GP-Refiner, and RACER? If direct comparisons are impractical due to differences in prediction targets or computational overhead, could the authors explain these limitations and clarify the advantages of the proposed mechanisms over these approaches?
  • In Table 3, adding state feedback increases the relative velocity prediction error from 0.1181 to 0.1262 and slightly decreases PSNR from 24.41 to 24.39, while improving LPIPS from 0.1788 to 0.1603. Could the authors elaborate on why state feedback improves perceptual fidelity despite worsening the measured velocity prediction accuracy? Additionally, could the authors provide separate ablations for clean-state feedback and corrected gain adaptation to clarify their respective contributions?
Reviewer_2DkmOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addersses the problem that, in training-free acceleration, step-skipping forecasts tend to over-extrapolate over long gaps and push the trajectory off course. It proposes to call the model only at a fixed small set of steps and to predict the velocity at the remaining steps by “trend + curvature” extrapolation from the last three velocities, constraining the prediction with order-specific gains, horizon-dependent damping and magnitude caps, together with residual calibration and state feedback. Acceleration is demonstrated on multiple base models. However, the method lacks an overall design rationale and is largely engineering-driven; it can be viewed as adding a set of anti-over-extrapolation regularizers to TaylorSeer.

Critical strengths

(1) The evaluation is thorough and the results are credible.

(2) The experiments are candid: the authors report unfavorable results themselves, which provides a valuable empirical insight for this line of work.

Critical weaknesses

(1) The method is a stack of empirical rules with more than a dozen hand-tuned parameters, and the paper lacks a design rationale and a description of the tuning procedure.

(2) Key comparisons are missing: results of the native scheduler run directly with few steps at the same NFE are not reported, and there is no comparison with MeanCache and FastFlow, which the paper cites as the closest works.

Additional comments:

In essence, the current paper can be viewed as adding a set of anti-over-extrapolation regularizers to TaylorSeer. I suggest rethinking the positioning of the paper, for example turning it from a “method paper” into a “finding + principle” paper. For instance: in training-free cache-based acceleration, the main cause of failure is not that the forecast is insufficiently accurate, but that it overshoots; controlling the forecast magnitude matters more than designing a more complex predictor.

Questions for authors

pls see weaknesses

Reviewer_5khiOpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to solve the high inference cost of flow-based image and video generation caused by repeated denoiser evaluations. Specifically, the authors propose a structured velocity-change predictor with separate trend and curvature controls to constrain sparse-anchor extrapolation, and residual calibration and bounded state feedback to correct forecasts while retaining the native scheduler. Experiments on DrawBench200 and VBench across four backbones report faster inference and improved PSNR, SSIM, and LPIPS over Spectrum at the main video operating point, while image results and more aggressive acceleration settings exhibit metric-dependent trade-offs.

Critical strengths
  1. The studied problem is interesting and important. Reducing repeated evaluations of large generative models without retraining addresses a substantial computational bottleneck.

  2. Separately scaling and bounding the trend and curvature is feasible to limit extrapolation increments.

  3. The evaluation setting is convincing. Multi-faceted metrics are used.

Critical weaknesses

It is essential to clearly show the function of different components.

In the evaluation, it is better to separate ablation studies and sensitivity analysis to justify the complexity of the correction process.

In some metrics, the method proposed in this paper is not obviously better, which should be clarified.

Questions for authors

see above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.