acceptodds
Withdrawn from ICLR 2027

STIR-FM: Segment-wise Trajectory Interpolation Refinement for Flow Matching

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Flow-matching models require multiple sequential model evaluations for text-to-image generation. We propose STIR-FM, a training-free post-sampling framework that refines a cached few-step trajectory by inserting model evaluations into intervals selected using velocity variation and propagating local velocity corrections. STIR-FM scores each interval by the velocity change across it, inserts an evaluation at the midpoint of the highest-scoring interval, and corrects the cached velocities downstream from a coordinate-wise spatial slope estimated from three evaluated reference pairs. An initial trajectory costing model evaluations and inserted evaluations has a total cost of NFEs, with only new evaluations needed when the initial trajectory is already available. Fixed-budget experiments on FLUX.1-dev and Stable Diffusion 3/3.5 show improved mean ImageReward over the initial trajectories; at matched total NFE, the refined means approach or exceed direct sampling once the initial trajectory is moderately resolved. A separate STORK experiment refines only samples falling below their per-sample STORK-12 CLIP reference and selects the highest-CLIP output in each history, reducing counted generator NFEs by % while mean held-out ImageReward remains numerically close to that of fixed STORK-12.

Primary area:
generative models
Keywords:
Flow Matching, Post-Sampling Refinement, Trajectory Refinement, Training-Free Inference, Text-to-Image Generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_a56qOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a method for refining an already-sampled Euler trajectory of a flow matching model by selectively adding additional function evaluations to improve the results. The method works by (1) scoring the error by looking for velocities from adjacent steps with large changes (2) evaluating those top-M intervals at their midpoint and (3) inserting a "correction" to the trajectory given this additional evaluation. The paper provides evaluations on different models (Flux.1-dev, SD3, SD3.5) by comparing the refined trajectories compared to their original trajectories on ImageReward and CLIP.

Critical strengths

The interval scoring is an interesting idea about where extra evaluations can help. In the placement ablation of Table 4, the proposed velocity-based selection beats random interval and uniform-time insertion baselines. For long trajectories (a>13 steps), refinement of b steps outperforms using a+b steps uniformly to begin with (although this does not hold for other values of a+b).

Critical weaknesses

First, I am unsure about the practicality of the problem and chosen problem setting, and the experiments do not address this. In what settings would one have a whole trajectory saved off and wish to refine it, vs. simply devoting those resources towards using those steps to sample to begin with? This paper never describes a situation in which this would be used. Furthermore, because this setting is not justified, the post-hoc (rather than online) application is an unnecessary burden. That is, why would one not use an online sampler and adaptively adjust the steps, and then not worry about having to propagate changes to already-sampled future steps? The paper does not address this baseline.

My second critical concern is regarding the results, both qualitatively and quantitatively.

  • Qualitatively, even though this paper is about text to image generation, there is a single visual result (Figure 2) and then no visual results until page 36 in the supplementary material. Looking at those images, I see almost no difference at all in their quality.
  • Quantitatively, most of the results appear to be negative. That is, using a+b (a original sampling; b refinement) typically underperforms simply using a+b sampling steps to begin with, for many values. This is not well-presented but can be read from Tables 1 & 2 on the diagonals. Figure 4 plots only a>=14 for FLUX, omitting the range where refinement loses. Direct sampling beyond 18 steps is never reported, so most points in Figure 4 have no equal-NFE comparison at all.
  • There is a lack of baselines. The tables compare just to Euler sampling as the baseline, without any adaptive sampler, such as DPM-Solver which the paper cites. It also takes effort for the reader to get the proper baseline of a+b Euler steps, as referenced in the previous bullet point.

Taken together, with a questionable problem setting along with what seem to be negative results. I believe this paper is currently not ready for publication.

Questions for authors

Please clarify the problem setting that is chosen (why a post-hoc method). What actually changes in the images after refinement that make it higher quality? Please tabulate the equal-NFE difference (a+b original sampling vs a sampling+b refinement) explicitly for direct reporting.

1 comment

Thank you for the review. We agree that equal-NFE comparisons should be clearer and that qualitative evaluations and baselines should be strengthened. However, these limitations do not establish that the problem is unjustified or that the results are generally negative.

Our method addresses refining an existing cached trajectory with additional computation, which differs from sampling from scratch or adapting steps online. The review itself acknowledges improvements in some equal-NFE settings and the effectiveness of interval selection. Failures at low initial step counts delimit applicability rather than negate those findings. Moreover, Table 3 includes STIR+STORK experiments, so the characterization of the evaluation as “Euler-only” is inaccurate. Although reference-based stopping limits this evidence, it should not be overlooked.

We therefore ask that presentation weaknesses, incomplete evidence, and methodological invalidity be assessed separately rather than conflated.

Reviewer_8mcdOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a novel training-free fast sampling method for flow-matching models. It utilizes the information within a trajectory that's already generated, and conducts interpolations of velocities in between existing velocities for finer approximations. Multi-layer velocity refinements are conducted for better generation quality. Experiments on the FLUX.1-dev and Stable Diffusion 3/3.5 shows the performance of the proposed method.

Critical strengths

Here are some strengths:

  1. The method description is clear and easy to follow.
  2. The experiments are conducted comprehensively across different approaches for velocity refinements.
Critical weaknesses

See the questions below.

Questions for authors

Here are some weaknesses:

  1. The baseline methods are only Euler and STORK; some other baseline comparisons like DPM-Solver++ would be beneficial, along with the proposed method applied to them.
  2. The CLIP score for the FLUX.1-dev is not so different across different methods; some other metrics that would differentiate the image quality more clearly might be more helpful for demonstrating the method's ability.
1 comment

Thank you for your thoughtful review and for recognizing the clarity of our presentation and the breadth of our experiments. We appreciate your constructive suggestions on additional solver baselines and evaluation metrics, which will help us strengthen the work. We also hope to have opportunities for further discussion and exchange of ideas in the future.

Reviewer_BMfBOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a training-free post-sampling framework for image generation using flow matching models. It finds the most error-prone intervals in the denising process of the original generation trajectory, and for those intervals, it performs a recomputation of velocity for the mid-point of that interval that makes the process smoother as compared to the original generation for the full trajectory with a small number of additional model evaluations. Experiments over Flux and Stable Diffusion backbones support the hypothesis of the paper.

Critical strengths
  1. The paper considers an important problem of efficiency and quality of flow image generation models.

  2. STIR-FM is fully training free and unsupervised. Its separation of interval selection, midpoint evaluation, and downstream velocity correction makes the framework potentially applicable to different flow-matching architectures and numerical solvers.

  3. The STIR+STORK experiment demonstrates conditional refinement, reducing counted generator evaluations by 9.27% while obtaining a numerically similar mean held-out ImageReward to fixed STORK-12.

Critical weaknesses
  1. The paper misses some very important and relevant baselines, that target the same problem and are quite relevant to this work, the paper neither compares with them nor cites them, which seems like a lack of literature review [1][2][3].

  2. The hypothesis that the interval partition is uniform is wrong. These intervals are already partitioned in such a way that given a budget the partition obtained performs significantly good by choosing the best partition from the interval. Hence the paper is only incremental in its work and the usefulness of the method is very thin.

  3. A central contribution is the coordinate-wise spatial slope estimator used to update cached velocities at displaced states. However, the experiments do not directly compare the full method against midpoint insertion without velocity correction, correction-free downstream propagation, or alternative inexpensive velocity approximations. Consequently, it is unclear how much of the observed improvement comes from adaptive insertion versus the proposed correction mechanism.

  4. While STIR-FM generally improves upon its initial trajectory, this is not sufficient to establish superior sampling efficiency. At matched total NFE, its performance is competitive mainly when the initial trajectory is sufficiently resolved. For example, FLUX.1-dev with 12+1 evaluations obtains ImageReward 0.838, compared with 0.847 for direct 13-step sampling. Even in the favorable regimes, some improvements are small and may not be statistically meaningful.

[1] https://proceedings.iclr.cc/paper_files/paper/2026/file/93e06b762457778f3e9832b9c90d5ad2-Paper-Conference.pdf [ICLR 2026]

[2] https://openreview.net/pdf?id=wWkyL8D9xd [ICLR 2026]

[3] https://openaccess.thecvf.com/content/CVPR2025/papers/Liu_Timestep_Embedding_Tells_Its_Time_to_Cache_for_Video_Diffusion_CVPR_2025_paper.pdf [CVPR 2025]

Questions for authors

See weaknesses.

1 comment

Thank you for the review. We respectfully disagree with several conclusions, particularly the conflation of related acceleration methods with our problem setting and the inference that optimized timestep schedules make trajectory refinement unnecessary.

1. Related work does not establish equivalence of the problems or methods.
FlowCast forecasts trajectories and verifies them in parallel, FastFlow skips evaluations through temporal velocity extrapolation, and TeaCache reuses cached computations during sampling. STIR-FM instead revisits an already-generated trajectory, allocates additional evaluations to selected intervals, and corrects downstream cached velocities for changes in state. Temporal extrapolation and spatial-displacement correction at a fixed timestep approximate different quantities. Missing these comparisons is an evaluation limitation; it does not, by itself, establish that our mechanism is merely incremental.

2. Optimized initial schedules do not make trajectory-specific refinement redundant.
Align Your Steps [1] demonstrates the benefits of optimizing timestep placement. More directly, OSS [2] and SharpEuler [3] improve sampling schedules for FLUX, while INDIS [4] demonstrates that globally optimized schedules can still benefit from instance-specific adaptation. These results undermine the inference that existing timestep partitions leave little value in further refinement. STIR-FM does not require uniform initial timesteps; it addresses where to allocate additional evaluations on an already-generated trajectory and how to propagate corrections using cached information.

3. Interval selection and velocity correction are distinct experimental questions.
Our placement ablation holds correction and propagation fixed, isolating the value of interval selection. We acknowledge that it does not independently quantify the correction mechanism’s contribution. A dedicated ablation would strengthen that evidence. Its absence, however, leaves an attribution question unresolved rather than demonstrating that the correction is ineffective or redundant.

4. A matched-NFE counterexample limits the scope of the gains but does not invalidate the refinement setting.
The FLUX.1-dev result of 0.838 for 12+1 versus 0.847 for direct 13-step sampling shows that refinement is not universally preferable when the entire budget is available beforehand. We do not dispute this. Improving an existing sample, however, differs from choosing an initial sampling budget: restarting an already-completed generation incurs additional computation. End-to-end and incremental-cost comparisons answer different questions.

As your review acknowledges, conditional STIR+STORK reduces counted generator evaluations by 9.27% on the reported 1,000-sample evaluation while obtaining a numerically similar mean held-out ImageReward. This is a bounded empirical result, not a claim of statistical equivalence or universal superiority; reference-generation and scoring costs remain separate.

We acknowledge the need for broader baselines, correction-specific ablations, and uncertainty estimates. Nevertheless, these limitations do not establish that online acceleration methods already address post-sampling refinement or that optimized initial schedules eliminate its potential value.

References:

[1] Align Your Steps: Optimizing Sampling Schedules in Diffusion Models, 2024.
[2] Optimal Stepsize for Diffusion Sampling, 2025.
[3] Sharpen Your Flow: Sharpness-Aware Sampling for Flow Matching, 2026.
[4] Few-Step Diffusion Sampling Through Instance-Aware Discretizations, 2026.

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

Related papers

Discussion (0)

Sign in to comment.