acceptodds
Under review as a conference paper at ICLR 2027

Lightweight Spectral Statistical Correction for Fast Sampling in Flow Models

Abstract

Diffusion and flow matching models achieve remarkable quality in various generation tasks. However, solving the flow ODE to a good precision typically requires many network evaluations, leading to high latencies. Multiple lightweight methods have been proposed for fast sampling with pretrained models. However, the vast majority of them focus on minimizing the truncation error of the underlying ODE solver, and do not attempt to directly control the distributional shift caused by taking large discretization steps. In this paper, we demonstrate that adding a simple fast analytical statistical correction after each solver step can significantly improve the quality of the generated images. Specifically, we present Spectral Statistical Correction (SSC) - a linear transformation applied independently to each DCT coefficient of the latent state so as to optimally correct its distribution. SSC can be learned from a small number of images without requiring backpropagation through the model. Furthermore, it can be applied on top of other fast sampling methods. As we show, SSC consistently leads to substantial improvements even over strong samplers, allowing to obtain high quality results with as little as 5 NFEs with modern flow models.

Primary area:
generative models
Keywords:
computer vision, generative modeling, diffusion models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_4yiGOpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Spectral Statistical Correction (SSC), a framework for improving few-step sampling from pretrained flow and diffusion models without updating model weights. After each solver step, SSC applies a per-coefficient affine correction in the DCT domain to match teacher statistics, followed by an optional correction of cross-channel statistics in a learned low-dimensional basis. The corrections are estimated sequentially using teacher trajectories. Experiments on FLUX.1 and SD 1.5 demonstrate improvements across several sampling methods, particularly in Teacher FID.

Critical strengths

Simple and practical statistical correction. SSC addresses systematic statistical deviations introduced by few-step sampling through analytically estimated transformations. It requires neither model fine-tuning nor backpropagation through the generative model, making it a potentially useful alternative to more expensive acceleration approaches.

Convincing improvements for standard solvers. The results for Euler, DPM++ (2M), and iPNDM (2M) show substantial improvements in Teacher FID, supported by improvements in ImageReward and HPSv2.1 on FLUX.1. These comparisons provide useful evidence that statistical correction can complement standard numerical solvers.

Critical weaknesses
  • The combinations with BézierFlow and DyWeight contain a confounding change to the sampling procedure. Section 5.1 states that these methods use shifted timesteps for model input, but that the shifts are removed when combining them with SSC because they conflict with the correction. Consequently, the reported comparisons change both the timestep treatment and the presence of SSC. They therefore do not isolate the incremental contribution of SSC. At minimum, the paper should compare each original method with the same method after removing the input-time shift, both without and with SSC. If preserving the original shift while applying SSC is infeasible, the incompatibility should be explained precisely.

  • The offline computational cost is insufficiently characterized and compared with competing methods. SSC avoids backpropagation, but still requires substantial forward computation. For FLUX.1, generating 1,000 teacher trajectories with 35 steps entails approximately 35,000 network evaluations, before accounting for sequential student calibration, PCA, and covariance estimation. The reported inference timing does not establish the total computational advantage of the method. The paper should report calibration wall-clock time, GPU-hours, peak memory, and a breakdown of teacher generation, student calibration, and statistical estimation. These costs should be compared with the training or optimization costs of DyWeight and other learned acceleration methods under comparable hardware and generation settings.

  • Typo: “timetseps” (Algorithm 1)

Questions for authors

See weaknesses above.

2 comments

BézierFlow and DyWeight time shifts evaluation (Weakness 1)

For FLUX with 5 NFE the following tables summarizes the relevant results:

MethodT. FID
DyWeight w.o. time shifts20.56
DyWeight w. time shifts16.48
DyWeight+SSC w.o. time shifts14.62
DyWeight+SSC w. time shifts12.13
BézierFlow w.o. time shifts14.30
BézierFlow w. time shifts13.99
BézierFlow+SSC w.o. time shifts11.52
BézierFlow+SSC w. time shifts12.03
Removing the timestep shifting from BézierFlow and DyWeight degrades their performance. Therefore, the reported results of BézierFlow and DyWeight without SSC in Table 1 are with input time shifting. For 5 NFEs, removing input time shifting increases BézierFlow teacher FID from 13.99 to 14.30, and increases DyWeight teacher FID from 16.48 to 20.56. This validates SSC’s improvement when combined with these methods. When using SSC with input time shifts, BézierFlow+SSC teacher FID increases from 11.52 to 12.03, which is still below the 13.99 of the BézierFlow with time shifts. However, with time shifts DyWeight+SSC teacher FID decreases from 14.62 to 12.13, farther below the 16.48 of DyWeight with time shifts. This result is better than the one we report, and we will add it to the final version.

Calibration cost (Weakness 2)

We performed timing experiments for 5 steps generation, using two A100 GPUs. SSC calibration wall clock time was 0.95 hours, with 1.9 GPU hours. Peak memory allocation was 41.7 GiB. Teacher generation took 0.75 wall clock hours, while fitting took 0.19 wall clock hours. SSC computations (PCA, covariance and student correction) took merely tens of seconds. For timing DyWeight training, we run the official training example script with 5 solver steps. The released checkpoint might not fully correspond to these settings, as not enough metadata is saved. Wall clock time was 5.2 hours, with 10.4 GPU hours. Peak memory allocation was 41 GiB.

The official BézierFlow implementation lacks FLUX.1. Fitting BézierFlow on FLUX.1 with the SD 3.5 settings from the paper takes 0.14 wall clock hours, and 0.29 GPU hours. However, under the settings we used for training BézierFlow (bigger train and validation sets, more epochs), fitting took 4.6 wall clock hours, and 9.2 GPU hours, with peak memory allocation of 35.8 GiB. We used these settings as they achieved a substainialy lower validation error than the SD 3.5 settings, thus making the comparison to BézierFlow fairer.

Reviewer_4yiGOpenReview ↗

Response

Thank you for the detailed response and additional experiments. The time-shift ablations clarify SSC’s independent contribution, and the timing measurements address my concern about calibration costs. With these clarifications and additional evaluation, my main concerns are addressed and I will raise my score.

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

The paper proposes a lightweight correction for distributional distortions caused by few-step sampling. It sequentially calibrates student trajectories against a higher-step teacher using per-coefficient affine transformations in a DCT basis and cross-channel Gaussian transport in a learned low-dimensional basis. Experiments report improved teacher-distribution matching and image-quality metrics on FLUX.1 and Stable Diffusion 1.5.

Critical strengths
  1. The paper offers a simple, practical way to improve few-step generation with limited additional computation. Its main value is an economical correction that can complement existing sampling methods.
  2. The method converts teacher–student statistical discrepancies into explicit affine transformations, requiring neither fine-tuning the generative backbone nor additional denoiser evaluations during sampling. The DCT and low-dimensional covariance corrections provide a computationally economical implementation of the proposed distributional perspective.
Critical weaknesses
  1. The experiments do not establish the added value over simpler or closely related statistical corrections. Section 3 discusses prior spectral corrections, but §5.2 compares primarily against solvers and schedule-optimization methods. The paper also lacks a matched global, per-channel, or coarse-frequency affine-calibration baseline. Consequently, it remains unclear how much benefit requires the proposed frequency-resolved construction rather than simpler contrast/statistics correction. This is a central gap in establishing the method’s incremental contribution.
Questions for authors
  1. At matched calibration data and NFE, how much improvement remains over a per-channel or coarse-frequency affine correction? Does the frequency-resolved design provide a measurable advantage beyond global contrast adjustment?
  2. How does SSC compare with a representative spectral-correction method discussed in §3 under the same backbone, solver, and sampling budget?
  3. For BézierFlow and DyWeight, what are the results for the original configuration, the configuration without input-time shifts, and that same configuration with SSC? Which part of the reported improvement is attributable to removing the shifts?
1 comment

Evaluation of simple corrections (Weakness 1, Question 1)

We evaluated two simple corrections: a per channel variance correction, and a 20 bins DCT correction. We will add these experiments to the final version of the paper. For the variance correction, we remove the DC, correct the variance, and then add the DC back. The DCT spectrum was binned by radial frequency. For each frequency bin, a single correction was estimated, and applied to each coefficient in the bin, independently for each channel. For FLUX.1 with the 5 NFE Euler solver the variance correction achieved teacher FID of 14.77 (FID measure w.r.t images generated by a high NFE teacher), while the binned DCT correction achieved teacher FID of 10.71. The Euler solver without correction achieves 17.64, while Euler+SSC achieves 9.48. These results show that a spectral correction is necessary, but the difference between a binned DCT and full DCT is not big, due to the smoothness of the DCT spectrum (on the ablation set, pure DCT correction achieved a teacher FID of 10.23, see Appendix A.4). SSC outperforms the binned DCT correction, due to the added channel covariance correction (see ablation in appendix A.4).

Evaluation of cited spectral corrections (Weakness 1, Question 2)

The spectral methods discussed in §3 are W++ of Yu and Zhan (2025) and SPA of Kobayashi et al. (2026). We implemented both methods with FLUX.1. Since our main metric is teacher FID we adapted them to use generated images with a 35 NFE teacher instead of real images. For W++ we performed parameter search similar to the original method. For SPA, we found the correction strength using cross validation. With 5 NFEs, W++ achieved teacher FID of 14.40, and SPA achieved teacher FID of 13.02. We will add W++ and SPA for 7 NFEs to the final version. SSC significantly outperforms these methods, with a teacher FID of 9.48. We note that both of these methods were presented for exposure bias mitigation, and not for fast sampling.

BézierFlow and DyWeight time shifts evaluation (Weakness 1, Question 3)

For FLUX with 5 NFE the following tables summarizes the relevant results:

MethodT. FID
DyWeight w.o. time shifts20.56
DyWeight w. time shifts16.48
DyWeight+SSC w.o. time shifts14.62
DyWeight+SSC w. time shifts12.13
BézierFlow w.o. time shifts14.30
BézierFlow w. time shifts13.99
BézierFlow+SSC w.o. time shifts11.52
BézierFlow+SSC w. time shifts12.03

The results of DyWeight and BézierFlow reported in the paper are with input time shifts. For the Euler solver with 5 NFEs, removing the input time shifts increases the teacher FID of the original BézierFlow from 13.99 to 14.30, and the teacher FID of the original DyWeight from 16.48 to 20.56. Therefore, the improvement of SSC without input time shifts can be attributed to SSC alone. When adding input time shifts to BézierFlow+SSC, the teacher FID increases from 11.52 to 12.03, which is still below the 13.99 of the original BézierFlow with time shifts. Adding input time shifts to DyWeight+SSC decreases the teacher FID from 14.62 to 12.13, which is significantly below the 16.48 of the original DyWeight with time shifts, and also better than the 14.62 of DyWeight+SSC without input time shift, which we report in the paper. For DyWeight+SSC this result is better than the one we report in the paper, and we will add it to the final version.

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

This paper proposes Spectral Statistical Correction (SSC), a lightweight method for improving few-step sampling with pretrained flow and diffusion models. After each solver step, SSC applies an affine transformation to individual latent DCT coefficients to match their means and variances to a many-step teacher. An additional learned-basis correction captures selected cross-channel correlations. The corrections are estimated sequentially from 1,000 trajectories without updating the generative model. Experiments evaluate SSC across several samplers on FLUX.1-dev and Stable Diffusion 1.5.

Critical strengths

The method is simple and motivated by observable differences in latent contrast and spectral power between coarse and fine sampling. Sequential estimation accounts for the influence of earlier corrections on later states. The reported improvements are substantial across several solvers: at five NFE, SSC reduces teacher FID from 13.10 to 7.51 for DPM++ and from 15.36 to 6.99 for iPNDM, alongside improved preference scores. Achieving these gains without model fine-tuning or additional denoiser evaluations makes the approach practically appealing.

Critical weaknesses
  1. The generalization and sample requirements of the calibration procedure are insufficiently evaluated. The paper uses 1,000 trajectories with different prompts, but does not clearly specify their source, composition, or separation from evaluation prompts. Since SSC applies one prompt-independent correction, its effectiveness may depend on how closely deployment prompts resemble the calibration distribution. Results across calibration-set sizes, independently sampled calibration sets, and different prompt domains would help establish whether the method is broadly reusable rather than specific to the tested setting.

  2. The statistical motivation also leaves the scope of the correction unclear. Few-step errors can involve missing objects, incorrect counts, or substantially different layouts, rather than only contrast and spectral power. Gaussianity of individual DCT coefficients does not imply that matching their marginal moments corrects these semantic and structural discrepancies. Subsequent nonlinear denoising could turn intermediate statistical corrections into structural improvements, but the current evaluation does not distinguish this from appearance restoration. A small learned corrector, such as a shallow encoder–decoder or residual CNN trained on the same trajectories, would also be a relevant comparison. Such a model need not update the generator and could capture state-dependent errors beyond a fixed affine transformation.

Questions for authors
  1. How were the 1,000 calibration prompts selected, and are their prompts and seeds disjoint from evaluation? Please summarize any existing results on calibration-set size, variability, and transfer to different prompt domains.

  2. Are comparisons available with a small learned corrector using the same calibration trajectories? How do quality, calibration cost, inference overhead, and generalization compare?

  3. Does SSC improve semantic and compositional accuracy, including object presence, counting, and spatial relationships, or are its gains primarily in contrast and texture? Please clarify the limits of the moment-matching approximation.

1 comment

Calibration prompts, calibration set size and transferability (Weakness 1, Question 1)

The calibration prompts were randomly sampled from the MS-COCO validation set, with IDs disjoint from the evaluation set.

To test the effect of calibration set size, we fitted SSC for FLUX.1 with the 5 NFE Euler solver using 200 and 500 prompts from our calibration set. We kept the same number of learned directions as the 1,000 prompt experiment reported in the paper (k=40). SSC calibrated on 200 prompts achieved teacher FID (FID measured w.r.t images generated by a high NFE solver) of 10.72, while SSC with 500 prompts achieved teacher FID of 9.73, compared to 9.48 using 1,000 prompts. We speculate that a separate cross validation for the number of directions, can improve these results. Our conclusion is that using 1,000 trajectories is a reasonable choice, balancing between compute and performance.

To test transferability to different prompt domains, we tested the 5 NFE FLUX.1 Euler SSC corrector trained on MS-COCO on the DiffusionDB dataset. It contains prompts users used to generate images, and is different in style and content from MS-COCO. We used 30,000 randomly sampled prompts from the dataset. While the 5 NFE Euler sampler achieved ImageReward of 0.69 and HPSv2.1 of 0.259, the Euler+SSC variant achieved ImageReward of 0.91 and HPSv2.1 of 0.288, showing significant improvement (for both metrics higher scores are better). This experiment shows evidence of transferability of the SSC correction, which we explain by the fact that we calibrate SSC on statistics of generated images. We will provide full teacher FID results in the final version.

Nonlinear learned corrector (Weakness 2, Question 2)

We trained a small residual CNN corrector for each step for 5 NFE Euler sampling with FLUX.1. Each CNN had 46k parameters. Similar to SSC, the CNNs were trained sequentially. After each student step, all the student states were gathered and the CNN was trained to match the student states and corresponding teacher solver states. The correction was applied, and the student states were advanced to the next timestep. To avoid overfitting, the number of training epochs was selected after each step using 5-fold cross validation. Apart from the first step, the CNN failed to reduce the MSE of the identity operation by more than 3%. We attribute this to the divergence of the student and teacher trajectories along the sampling paths. The CNN achieved a teacher FID of 49.73, accompanied by notable visual artifacts. This FID is significantly higher than the 17.64 teacher FID of the uncorrected Euler solver. We explain this result by the paired student-teacher training of the CNN, which stands in sharp contrast to the distributional calibration of SSC.

Evaluation of semantic and textural accuracy of SSC (Weakness 2, Question 3)

We evaluated SSC for FLUX.1 on GenEval with the Euler solver for 5 and 7 NFEs. GenEval measures how well a model is able to follow prompts, by analyzing objects' presence, count, spatial relations, color and color attribute binding. We followed the standard evaluation protocol of 553 prompts and 4 images per prompt.

For 5 steps, adding SSC to the Euler solver improves the overall score from 0.51 to 0.54 (p-value=0.0013). At a significance level of 0.05 (using a two-sided t-test) SSC improves colors, color attribution and two objects categories scores. There are no other categories with significant differences. For 7 steps, adding SSC to the Euler solver improves the overall score from 0.58 to 0.59 (p-value=0.21). At a significance level of 0.05, SSC improves only the color attribution category. There are no other categories with significant differences.

This shows that SSC mainly improves appearance related features, like color and texture. However, this is done in a semantically meaningful way, as indicated by the color attribution category, which measures the correct attribution of colors to objects.

Full results for 5 steps:

Metric5 NFE Euler (%)5 NFE Euler+SSC (%)Difference (pp)Two-sided p-value
single_object94.6995.620.940.369
two_object54.2960.105.810.0298
counting61.5655.94-5.620.063
colors63.8370.746.910.0283
position14.0014.750.750.633
color_attr15.7526.5010.750.000139
Overall GenEval50.6953.943.260.00129

Full results for 7 steps:

Metric7 NFE Euler (%)7 NFE Euler+SSC (%)Difference (pp)Two-sided p-value
single_object96.5697.811.250.159
two_object69.4468.94-0.510.829
counting63.4459.69-3.750.208
colors75.0076.601.600.521
position18.2518.250.001
color_attr26.2534.508.250.00175
Overall GenEval58.1659.301.140.213

While moment matching fixes only some aspects of the distribution, it is applied before subsequent model evaluations (excluding the last step), which add powerful nonlinearity.

Comments on the paper

2 comments

As the discussion period ends tomorrow, we would like to ask whether the reviewers have any important remaining concerns or questions. We believe we have thoroughly addressed the feedback raised so far and would welcome the opportunity to provide any further clarification or respond to additional comments or concerns.

AuthorsOpenReview ↗

New Paper Version

We uploaded a new version of the paper.

For the main FLUX.1 experiments (Table 1), we changed the BézierFlow and DyWeight combined with SSC experiments, to be with time shifts. We added a new section to the appendix (A.5), that contains analysis of BézierFlow and DyWeight without time shifts, both with and without SSC.

Additionally, we added to Table 1 an evaluation of the alternative spectral corrections W++ and SPA for both 5 and 7 NFEs.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.