acceptodds
Under review as a conference paper at ICLR 2027

SeAL: Accelerating Flow Matching through Segment Advancement with Shared Lookahead

Abstract

Flow-matching models generate high-quality images and videos, but repeated neural network evaluations make sampling expensive. Skipping evaluations reduces this cost by extending an available velocity estimate over a longer span. However, local velocity agreement alone does not determine a suitable span, and checking each candidate endpoint adds costly model calls. We introduce ReAL, a training-free sampler that selects how far to advance using one shared lookahead. Our key insight is that the discrepancy between uncorrected and lookahead-corrected endpoint proposals can be computed directly from the observed velocity mismatch and the candidate span beyond the lookahead. This relation provides a span-dependent selection criterion without additional endpoint evaluations. The same lookahead selects the longest passing candidate span, corrects the accepted update, and supplies its velocity as the next starting estimate. After initialization, each regular iteration requires only one fresh evaluation. ReAL uses the pretrained velocity output and original noise schedule, with no additional training or access to internal features. Experiments cover four image-generation backbones, video generation, and image editing. ReAL achieves 4.91× measured speedup on FLUX.1-dev while retaining 97.0% of dense mean ImageReward. On HunyuanVideo, it achieves a 5.49× speedup while maintaining a VBench score close to that of dense sampling.

Primary area:
generative models
Keywords:
Flow Matching, Efficient Sampling, Training-Free Acceleration, Adaptive Sampling, Generative Models, Diffusion Models

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_23M8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies training-free acceleration for flow-matching models. The main question is how far the sampler can advance using a single velocity evaluation. Simply reusing the same velocity over a long interval can significantly hurt generation quality. To address this, the authors propose ReAL, which performs a one-step lookahead and compares the current velocity with the velocity at the next step to estimate whether the local trajectory is sufficiently straight. This signal is then used to select the advancement length, correct the resulting endpoint, and provide the velocity for the next iteration. Experiments on four text-to-image models, HunyuanVideo, and Step1X-Edit show substantial speedups while largely preserving the quality of dense sampling.

Critical strengths
  1. The main idea is simple and well motivated. ReAL uses one nearby lookahead evaluation to check whether the current velocity is still reliable and then adapts how far the sampler advances. The same lookahead is also reused for endpoint correction and the next iteration, making the added computation worthwhile.
  2. The experiments are comprehensive, and the method shows consistent effectiveness across multiple base generation models.
Critical weaknesses
  1. The method relies on several hand-designed choices, including the maximum span, the candidate backoff rule, and the local and segment thresholds. Since these settings need to be adjusted for different models and tasks, they introduce additional calibration costs in practice. They may also become unreliable outside the calibrated settings, limiting the method’s extensibility.
  2. The method makes acceleration decisions mainly based on local information from a one-step lookahead. This may limit how much it can speed up the model without losing quality. It may therefore be less effective than training-based acceleration methods, limiting its use in practice.
Questions for authors
  1. The paper notes that the maximum span, local threshold, and segment threshold are calibrated for each model and sampling setting based on generation quality and latency. What do the authors consider an efficient way to select these parameters for a new model, and what criteria can be used to determine whether the chosen configuration is best?
  2. Although ReAL is training-free, could the authors also compare it with one or two strong training-based acceleration methods, if comparable settings are available? This would allow for a better evaluation of ReAL’s capabilities.
Reviewer_7zFWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the difficulty of deciding how far to jump when accelerating flow-matching sampling, where verifying each candidate endpoint separately is costly. It performs a single lookahead evaluation per iteration: the discrepancy between the fresh velocity at the lookahead point and the current velocity determines the jump length, the fresh velocity corrects the jump endpoint, and it is carried to the next iteration. On FLUX, HunyuanVideo and Step1X-Edit, it achieves about 4.9× / 5.5× speedup with good quality retention and essentially no extra memory. However, the jump rule is essentially adaptive step-size control based on a local error estimate, yet the paper only compares with caching methods, lacks comparisons with solver-based methods, and provides no error guarantee.

Critical strengths

(1) The speedup is substantial without increasing memory.

(2) The experiments are solid, and the ablations clearly separate the role of each component.

Critical weaknesses

(1) The proposed jump rule is essentially adaptive step-size control based on a local error estimate, and carrying the lookahead velocity to the next iteration resembles the FSAL property of the Dormand–Prince method [1]. Please add the relevant citations.

(2) Strictly speaking, the method caches no intermediate features and is a sampler rather than a caching method, yet the experimental tables mainly compare against caching methods and lack comparisons with samplers. I suggest adding adaptive step-size solvers and multistep solvers that use one model evaluation per step (e.g., DPM-Solver++ and UniPC).

(3) There is no error guarantee: the paper gives no bound on the error introduced by accepted long jumps, so whether the jump rule is safe relies entirely on empirical tuning.

[1] J. R. Dormand, P. J. Prince. A family of embedded Runge–Kutta formulae. Journal of Computational and Applied Mathematics, 6(1): 19–26, 1980.

Questions for authors

pls see weaknesses.

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

This paper proposes ReAL, a training-free acceleration method for flow-matching inference based on shared lookahead. At each iteration, ReAL uses a carried velocity to construct a one-step lookahead state, evaluates the velocity field once at that state, and reuses the resulting velocity for segment selection, endpoint correction, and initialization of the next iteration. Candidate segment lengths are evaluated using a local velocity discrepancy and a span-scaled discrepancy, followed by hierarchical backoff and velocity carry. The method is evaluated on image generation, video generation, and image editing tasks.

Critical strengths
  1. The shared-lookahead mechanism is simple and computationally attractive. A single fresh velocity evaluation is reused for segment selection, endpoint correction, and velocity carry, avoiding separate model evaluations for different candidate endpoints.
  2. The experimental evaluation is broad. The paper studies image generation, video generation, and image editing, and includes ablations of local verification, span verification, endpoint correction, and cross-segment reuse.
Critical weaknesses
  1. Limited novelty and insufficient motivation. ReAL appears to be a relatively straightforward refinement of direct step skipping: it uses lookahead information to adapt the skip length and decide whether a skipped segment should be accepted. Although this design may improve the accuracy–efficiency tradeoff, the paper does not identify a sufficiently new principle or technical insight beyond making the skip decision adaptive. The motivation for the specific shared-lookahead design is also underdeveloped, making it difficult to view ReAL as a substantial methodological contribution rather than an incremental engineering improvement.
  2. The core acceptance criterion is heuristic and lacks a clear error justification. ReAL uses the local velocity discrepancy multiplied by the proposed advancement span to decide whether a segment is safe. However, the paper does not establish a quantitative relationship between this score and the final sampling error. The carried velocity is also evaluated at an earlier lookahead state rather than at the final accepted endpoint, so carry error may accumulate over multiple accepted segments. More generally, the paper does not clearly justify why one nearby lookahead observation should be sufficient to support a longer segment.
  3. The baseline comparison is incomplete and inconsistent across tasks. VDE and ZEUS are included in the HunyuanVideo comparison but are absent from the main FLUX image-generation comparison. The qualitative results also appear inconsistent with the reported quantitative performance; for example, VDE appears visually worse than ZEUS even though it achieves better quantitative metrics. The authors should explain this discrepancy and verify whether VDE and ZEUS were reproduced correctly according to their original papers.
Questions for authors
  1. What evidence supports using the span-scaled local discrepancy as a proxy for final sampling error? Can the authors report the relationship between the proposed score and the actual dense-versus-ReAL error for different segment lengths?
  2. How are the local threshold, segment threshold, and segment cap selected for each backbone? Were these parameters selected on an independent validation set or on the evaluation prompts?
  3. What are the VDE parameters reported in Appendix G? They appear to be inconsistent with the parameters described in the original VDE paper. Please clarify their meaning and configuration.
  4. Why are VDE and ZEUS included in the HunyuanVideo comparison but omitted from the main FLUX comparison? Also, why do the qualitative examples appear to rank VDE below ZEUS when the quantitative results may suggest the opposite?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.