acceptodds
Under review as a conference paper at ICLR 2027

SNAP: ONE-STEP POSTERIOR SAMPLING FOR NOISY INVERSE PROBLEMS

10 views

Abstract

Diffusion and flow-matching models can produce high-quality posterior samples for inverse problems, but typically require tens to thousands of network evaluations per draw. MeanFlow enables one-step generation, yet applying it to inverse problems leaves no intermediate steps at which to enforce measurement consistency. We introduce SNaP, a *one-step* MeanFlow posterior sampler for linear inverse problems with Gaussian noise. Its central innovation is a measurement-adapted source: a Gaussian distribution whose mean and anisotropic covariance are determined by the measurement operator, observation, and noise level. The source anchors well-measured directions while preserving variation where the measurements are weak or uninformative. We show that the exact conditional flow transports this source to the true posterior. Across natural-image restoration and multi-coil MRI, SNaP produces diverse, high-quality samples with one network evaluation per draw, 30 to 2250 faster than iterative samplers.

Primary area:
generative models
Keywords:
posterior sampling, inverse problems, one-step generative models, flow matching

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_8Pw4OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SNaP, a one-step posterior sampling method for linear inverse problems. The authors derive a measurement-aware Gaussian noise distribution for linear forward operators. The mean is chosen as the optimal estimate that minimizes MSE with the ground truth image; the covariance is chosen such that uncertainty is suppressed in well-measured directions while remaining large in weakly measured and null-space directions. A conditional MeanFlow is then trained on (clean image, measurement) pairs, using the measurement-aware Gaussian as the noise prior rather than a standard Gaussian.

Extensive experiments show strong performance on image reconstruction tasks (inpainting, super-resolution, deblurring) across the CelebA and AFHQ-Cat datasets, as well as multi-coil fastMRI evaluation. SNaP outperforms/is competitive with many diffusion/flow baselines at only 1 NFE. Ablations help validate specific design choices such as the source distribution.

Critical strengths

1. The problem is timely and well-motivated. Enabling one-step generation for posterior sampling is novel and potentially practically useful. Empirical results across distinct benchmarks are convincing: SNaP shows strong performance at only 1 NFE. The image reconstruction calibration results, combined with the synthetic Gaussian mixture experiment, are evidence that SNaP captures the full posterior distribution well.

2. The idea of a measurement-aware noise prior that suppresses uncertainty in the well-defined directions, but encourages uncertainty in the poorly measured directions, under the linear operator is interesting and novel to my understanding. The authors do extensive theoretical analysis that justifies most of the methodological design choices.

Critical weaknesses

1. The method requires training a new conditional MeanFlow model for each forward operator, which is precisely the limitation that training-free methods like DPS avoid. This is understandable because the method trades off training time to retain one-step inference, which training-free methods lack, so I don't view this as a fatal weakness. But I believe that this still weakens the practical utility of the method.

2. The experiments don't include comparison against relevant consistency-based inverse problem solvers [1,2,3], nor are they discussed in the related work. These are important baselines because they also focus on using one/few-step generative models for inverse problems.

Questions for authors
  1. (Major Weakness 1) It's understandable that SNaP requires forward operator-specific training. But I'd like to see a demonstration of generalizability across the forward operator family: i.e., how does SNaP performs on box inpainting if it trains on a distribution of box masks, then is evaluated on a held-out evaluation distribution of masks? I believe that demonstration of such a generalization capability would significantly improve the method's practical utility.

  2. (Major Weakness 2) The main experiments are missing a comparison against another one/few-step inverse solver. Can at least one of [1–3] be evaluated on an image reconstruction or MRI benchmark to enable a fair comparison with SNaP? Such a baseline would help determine whether SNaP’s 1-NFE performance is compelling relative to other methods specifically designed for low-NFE inverse solving. The discussion and appendix comparisons with NullFlow and VFM are appreciated, but both don't appear as baselines in the main image reconstruction/MRI experiments. The related-work section should also discuss [1–3].

  3. (moderate Weakness 3) How do other simple choices of CC affect SNaP performance? How robust is SNaP to this design choice? Since the method derivation works for general CC, this is an important design choice that should be studied more in depth.

  4. (clarification) What does SNaP stand for?

Author response has satisfied my concerns. Raising score from 2 to 3.

[1] M. Gülle, J. Yun, Y. U. Alçalar, and M. Akçakaya. “PnP-CM: Consistency Models as Plug-and-Play Priors for Inverse Problems.”

[2] T. Garber and T. Tirer. “Zero-Shot Image Restoration Using Few-Step Guidance of Consistency Models (and Beyond).”

[3] J. Zhao, B. Song, and L. Shen. “CoSIGN: Few-Step Guidance of ConSIstency Model to Solve General Inverse Problems.”

4 comments

We thank the reviewer for the feedback. We address each point with new experiments: we train one model on a family of box masks and test it on unseen mask sizes (W1), compare SNaP with PnP-CM on CelebA and fastMRI and with PnP-CM and CM4IR on CelebA-HQ (W2), and evaluate four additional working covariances CC (W3). See our point-by-point responses below.

(Major Weakness 1) It's understandable that SNaP requires forward operator-specific training. But I'd like to see a demonstration of generalizability across the forward operator family : i.e., how does SNaP performs on box inpainting if it trains on a distribution of box masks, then is evaluated on a held-out evaluation distribution of masks? I believe that demonstration of such a generalization capability would significantly improve the method's practical utility.

Prompted by the reviewer's comment, we trained SNaP on CelebA box inpainting with a randomly placed 40×40 box for each training image. The method needs no change, since the source is built in closed form from whatever mask is given.

One model trained on the mask family generalizes to unseen box sizes, gaining 5.9–6.9 dB over the fixed-mask model with a single draw.

We test on 100 images with new random boxes at the training size and at two sizes never seen in training, and score the submitted fixed-mask model on the same masks (PSNR at M = 1 / M = 4).

Test masksFamily-trainedSubmitted (fixed mask)
Random 28×28 (unseen size)31.51 / 33.5625.44 / 26.92
Random 40×40 (training size)28.40 / 30.4921.52 / 23.34
Random 52×52 (unseen size)24.33 / 25.7618.48 / 20.20
Centred 40×40 (submitted mask)29.14 / 31.3130.73 / 32.78

The family-trained model is lower only on the centred mask the submitted model was trained for, by about 1.5 dB.

We have revised the paper to include this experiment in App. B.3 (Table 9). The Limitations now state that a new type of operator still needs a new model.

(Major Weakness 2) The main experiments are missing a comparison against another one/few-step inverse solver. Can at least one of [1–3] be evaluated on an image reconstruction or MRI benchmark to enable a fair comparison with SNaP? Such a baseline would help determine whether SNaP’s 1-NFE performance is compelling relative to other methods specifically designed for low-NFE inverse solving. The discussion and appendix comparisons with NullFlow and VFM are appreciated, but both don't appear as baselines in the main image reconstruction/MRI experiments. The related-work section should also discuss [1–3].

We compare SNaP with PnP-CM [1] and CM4IR [2] in two settings, and discuss [1–3] in the related work.

At matched 4 NFEs, SNaP outperforms PnP-CM on every metric in both settings and CM4IR on every metric except SR ×4 PSNR, where CM4IR is 0.37 dB higher. In our setting, even a single SNaP draw achieves higher PSNR than PnP-CM at 4 NFEs in seven of eight tasks.

1. PnP-CM adapted to our setting. We distill consistency models on our CelebA and fastMRI training splits and run PnP-CM at 4 NFE with our operators.

TaskPnP-CM (4 NFE)SNaP, M=1 (1 NFE)SNaP, M=4 (4 NFE)
CelebA 128 (PSNR / LPIPS)
Deblur33.60 / 0.04632.90 / 0.01834.98 / 0.022
SR ×229.26 / 0.06331.27 / 0.02333.27 / 0.024
Random inp. (70%)29.00 / 0.06331.98 / 0.01833.97 / 0.016
Box inp. (40×40)27.82 / 0.05330.73 / 0.02132.78 / 0.019
fastMRI brain (PSNR / SSIM)
×4, 20 dB30.50 / 0.73432.07 / 0.89032.68 / 0.905
×4, 30 dB30.60 / 0.72932.89 / 0.89933.80 / 0.920
×8, 20 dB25.13 / 0.58228.11 / 0.81929.07 / 0.844
×8, 30 dB25.14 / 0.58129.12 / 0.84529.89 / 0.865

At matched 4 NFEs, SNaP achieves higher PSNR and better perceptual or structural metrics across all eight tasks. Even at 1 NFE, SNaP achieves higher PSNR in seven of eight tasks.

2. SNaP adapted to PnP-CM's setting. To keep the comparison fair to PnP-CM, we train SNaP on its operators and training-set size, and run PnP-CM and CM4IR with PnP-CM's released CelebA-HQ 256 checkpoint, σy=0.05\sigma_y = 0.05 and 300 test images (our rerun reproduces PnP-CM's paper within 0.15 dB). PSNR / LPIPS-VGG.

TaskPnP-CM (4 NFE)CM4IR (4 NFE)SNaP, M=1 (1 NFE)SNaP, M=4 (4 NFE)
Deblur28.93 / 0.24928.30 / 0.29527.37 / 0.25429.41 / 0.217
SR ×427.27 / 0.28527.88 / 0.28925.41 / 0.32827.51 / 0.275
Inpainting (70%)29.08 / 0.20429.55 / 0.28028.28 / 0.24230.38 / 0.198

At matched 4 NFEs, SNaP achieves the best performance on five of six metrics, with the sole exception of SR ×4 PSNR, where CM4IR is 0.37 dB higher.

We have revised the paper to include these results (Tables 1 and 3, App. B.2) and a discussion of the related works [1–3] (Sec. 5).

(moderate Weakness 3) How do other simple choices of CC affect SNaP performance? How robust is SNaP to this design choice? Since the method derivation works for general CC, this is an important design choice that should be studied more in depth.

Prompted by the reviewer's comment, we train SNaP with sources built from several working covariances CC.

SNaP is robust to the choice of CC at a fixed mean variance. Across the covariances tested, PSNR varies by at most 0.25 dB and Δcal stays within [0.90, 1.03].

Besides τ2I\tau^2 I, we test two classes of working priors, all with the same mean variance τ2=0.152\tau^2 = 0.15^2 (CelebA deblurring).

  1. Per-pixel diagonal, with entries drawn uniformly at random.
  2. Colored, stationary covariances with power-law spectra Σ(ω)∝(1+∥ω∥2/ω02)−α\Sigma(\omega) \propto (1 + \lVert\omega\rVert^2/\omega_0^2)^{-\alpha}, normalized so that every pixel has variance τ2\tau^2.
CCPSNR M=1PSNR M=16LPIPS M=1Δcal (→1)
τ2I\tau^2 I (paper)32.9435.690.0191.03
Per-pixel diagonal, ci∼U(0.0352, 0.2092)c_i \sim U(0.035^2,\ 0.209^2)32.9535.500.0180.90
Per-pixel diagonal, ci∼U(0.1062, 0.1842)c_i \sim U(0.106^2,\ 0.184^2)32.9735.630.0180.96
Colored, α = 1.5, ω₀ = 232.7335.450.0200.96
Colored, α = 0.5, ω₀ = 1032.9035.700.0191.01

We have revised the paper to include this study in App. B.6 (Table 15).

Q1. What SNaP stands for.

SNaP stands for (S)ampling for (N)oisy inverse (P)roblems.

Thanks for the response. The new results have answered my concerns; I'm increasing my score from 2 to 3.

Thank you for considering our response and raising your score. If any concerns are still affecting your assessment, we’d be happy to address them.

Reviewer_cRZJOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SNaP, a one-step posterior sampler for noisy linear inverse problems based on training (per operator/task) a measurement-conditioned MeanFlow where the input distribution (source) is a Tikhonov estimation with noise covariance that depends on the observation model (larger variance in null space directions). Experiments show correlation between diversity of the one-step samples and the error of their average. With sufficient samples, the average is also competitive with alternative methods (mostly zero-shot methods).

Critical strengths
  • The design of the input is simple and motivated by linear MMSE estimation (with "isotropic working covariance" relaxation).

  • The proposed one-step sampling is empirically shown to yield better calibrated spread than alternative methods on CelebA 128x128.

Critical weaknesses
  • Missing related works:

The recent work [1] builds similar measurement-aware anisotropic Gaussian, covariance (τ−2I+σn−2A⊤A)−1(\tau^{-2}I+\sigma_n^{-2}A^\top A)^{-1} and an estimate as pivot, within a multi-step posterior sampler. While I view this work as concurrent, please discuss the relation between the works.

Few-step zero-shot solvers, e.g., CM4IR [2] and LATINO [3], reach strong results in about 4--8 NFEs without per-operator training, and are directly related to the efficiency claims. CM4IR's noise injection also applies to diffusion-based methods such as DDRM and DiffPIR at 4 NFEs. This literature should be discussed and compared against.

[1] "Exact Posterior Score Estimation for Solving Linear Inverse Problems", https://arxiv.org/abs/2606.17048

[2] "Zero-Shot Image Restoration Using Few-Step Guidance of Consistency Models (and Beyond)", CVPR 2025

[3] "LATINO-PRO: LAtent consisTency INverse sOlver with PRompt Optimization", ICCV 2025

  • Good distortion results require many draws. The best PSNR/SSIM results (Table 1) use 16--100 averaged draws (16--100 NFE), and with few draws (M <= 4) the proposed SNaP is outperformed in PSNR by the MSE regressor on every task where the regressor is reported, as well as by several zero-shot methods. The speed comparison (Table 4), however, is reported for a single draw. So, in multiple places in the paper, the one-NFE and acceleration framing obscures the distortion cost. Such statements (including in the abstract) should be toned-down and provide the bigger picture.

  • Proposition 2 holds for any non-degenerate Gaussian source, including the isotropic sources of Table 5, which nonetheless collapse in practice. So Proposition 2 does not support the specific source design. This should be emphasized in the text.

  • SNaP is trained on paired data for each operator, yet is compared mainly with zero-shot methods, and the reported speedups exclude training.
    Compare with other per-operator one/few-steps posterior samplers, such as recent bridge models. They have similar advantages (single/few steps + diversity rather than point estimate).

  • Partial source ablation. Table 5 compares only zero-centered isotropic sources with the full source, so the contributions of the center mym_y and of the anisotropic covariance WW cannot be separated. Please complete the source ablation (Table 5) with N(my,τ2I)\mathcal{N}(m_y,\tau^2 I) and N(0,τ2W)\mathcal{N}(0,\tau^2 W).

  • Does your architecture depend on y, A, sigma_n beyond its input? From the appendix it seems that yes, but this should be clarified in the main text and motivated (currently it seems that the motivation is only for setting the input). Also, ablate such conditioning to show how much the source itself contributes.

  • State (e.g., in the appendix) the hyperparameters that you use for the competitors. Also what prior models do you use for them? (Provide details/links).
    I also suggest that you examine SR on the common bicubic SR benchmark instead of on the uncommon and less physical setup that you use (with no kernel).

  • The discussion on how to set τ\tau should be provided in the main paper, together with sensitivity examination. In particular, report how different choices affect the calibration metric. Related to this, currently the calibration metric is reported only on CelebA. Since this seems to be a selling point of your approach, please report on the other datasets.

  • Since the method is operator-specific and not zero-shot, please provide analysis of how mismatch in the observation model (different kernel in A and under-estimated noise level sigma_n) affects its performance.

  • The AFHQ-Cat results show weaker performance of the proposed method.

Questions for authors

The questions appear above.

9 comments

We thank the reviewer for the feedback. Prompted by the reviewer's comments, we compare SNaP with CM4IR and PnP-CM at the same 4 NFE (W2), with the bridge model I2SB trained on the same pairs (W5), and with the baselines at matched inference compute (W3, W12). We also complete the source ablation (W6), ablate the network conditioning (W7), add the bicubic SR benchmark (W9), report calibration on every dataset and across τ\tau (W10), and test mismatched noise levels and kernels (W11). In the paper, we discuss EPS, CM4IR and LATINO-PRO (W1, W2), restate the one-step framing in the abstract (W3), clarify Proposition 2 (W4) and the conditioning (W7), move the choice of τ\tau to the main text (W10), add the baseline details (W8), and state per-operator training as a limitation.

Missing related works: The recent work [1] builds similar measurement-aware anisotropic Gaussian, covariance (τ−2I+σn−2A⊤A)−1(\tau^{-2} I + \sigma_n^{-2} A^\top A)^{-1} and an estimate as pivot, within a multi-step posterior sampler. While I view this work as concurrent, please discuss the relation between the works. [1] "Exact Posterior Score Estimation for Solving Linear Inverse Problems", https://arxiv.org/abs/2606.17048

Prompted by the reviewer's comment, we have added [1] to the related work (Sec. 5) and discuss how it differs from SNaP. Unlike [1], which repeatedly computes a schedule-dependent Gaussian for score estimation in multi-step diffusion, SNaP constructs a measurement-dependent Gaussian source once and uses it for one-step posterior sampling.

Few-step zero-shot solvers, e.g., CM4IR [2] and LATINO [3], reach strong results in about 4--8 NFEs without per-operator training, and are directly related to the efficiency claims. CM4IR's noise injection also applies to diffusion-based methods such as DDRM and DiffPIR at 4 NFEs. This literature should be discussed and compared against. [2] "Zero-Shot Image Restoration Using Few-Step Guidance of Consistency Models (and Beyond)", CVPR 2025 [3] "LATINO-PRO: LAtent consisTency INverse sOlver with PRompt Optimization", ICCV 2025

Prompted by the reviewer's comment, we compare SNaP with CM4IR [2] at the same 4 NFE, discuss [2, 3] in the related work, and state per-operator training as a limitation.

At matched 4 NFE, SNaP achieves the best performance on every metric except SR ×4 PSNR, where CM4IR is 0.37 dB higher.

CM4IR is published on LSUN and ImageNet only, so we run it in PnP-CM's setting with the released CelebA-HQ 256 consistency model (PSNR / LPIPS-VGG). Our reply to Reviewer 8Pw4 (W2) also compares SNaP with PnP-CM on CelebA and fastMRI.

NFESR ×4DeblurInpainting
CM4IR427.88 / 0.28928.30 / 0.29529.55 / 0.280
PnP-CM427.27 / 0.28528.93 / 0.24929.08 / 0.204
SNaP, M = 4427.51 / 0.27529.41 / 0.21730.38 / 0.198

Relation to LATINO-PRO [3]. LATINO-PRO uses a latent consistency model (SDXL) at 512–1024 px with a text prior and prompt optimization. A direct comparison with SNaP's 128/256 px pixel-space models would therefore involve substantially different model architectures, resolutions, and prior information. We explicitly acknowledge SNaP's pixel-space formulation as a limitation and identify extending its source construction to latent-space flow models as future work.

We have revised the paper to discuss [2, 3] in Sec. 5, including CM4IR's noise injection for DDRM and DiffPIR at 4 NFE, and report the comparison in App. B.2 (Table 8).

Good distortion results require many draws. The best PSNR/SSIM results (Table 1) use 16--100 averaged draws (16--100 NFE), and with few draws (M <= 4) the proposed SNaP is outperformed in PSNR by the MSE regressor on every task where the regressor is reported, as well as by several zero-shot methods. The speed comparison (Table 4), however, is reported for a single draw. So, in multiple places in the paper, the one-NFE and acceleration framing obscures the distortion cost. Such statements (including in the abstract) should be toned-down and provide the bigger picture.

Prompted by the reviewer's comment, we add the cost of M = 4, 16 and 100 draws to Table 4, compare at matched compute, and revise the abstract, introduction and conclusion accordingly.

One evaluation gives a posterior sample, averaging draws trades compute for distortion, and at matched compute SNaP outperforms the iterative samplers in PSNR and SSIM. The revised abstract states all three.

The abstract now reads as follows.

Across natural-image restoration and multi-coil MRI, SNaP draws a posterior sample in one network evaluation, making each sample draw at least 30×30\times faster than iterative samplers. Averaging a few draws trades compute for distortion, and at matched compute SNaP is competitive with iterative samplers.

1. Cost of M draws (Table 4). NFE and 1-batch wall-clock time per posterior sample. SNaP with M draws costs M evaluations.

CelebA, Gaussian deblurring

DPSOT-ODEFlowerDDRMSNaP (M=1)SNaP (M=4)SNaP (M=16)SNaP (M=100)
NFE ↓1000180100201416100
Time / sample (s) ↓27.0146.5752.6140.3600.0120.0480.1921.200
Slowdown vs. SNaP (M=1) ↓2251×548×218×30×–4×16×100×

fastMRI brain, multi-coil ×8

CSGMDAPSPnP-DMDPSSNaP (M=1)SNaP (M=4)SNaP (M=16)SNaP (M=100)
NFE ↓10001000100010001416100
Time / sample (s) ↓127.7896.2892.0276.070.2270.9083.63222.700
Slowdown vs. SNaP (M=1) ↓563×424×405×335×–4×16×100×

2. Matched compute. CelebA, PSNR / SSIM.

About 4 NFE

MethodNFEDeblurSRRandom inp.Box inp.
PnP-CM433.60 / 0.93429.26 / 0.87729.00 / 0.89727.82 / 0.913
SNaP, M = 1132.90 / 0.91931.27 / 0.90331.98 / 0.92830.73 / 0.934
SNaP, M = 4434.98 / 0.94733.27 / 0.93533.97 / 0.95132.78 / 0.954

About 20 NFE

MethodNFEDeblurSRRandom inp.
DDRM2035.02 / 0.94632.24 / 0.92732.40 / 0.942
SNaP, M = 161635.68 / 0.95433.95 / 0.94334.64 / 0.957

100 NFE and above

MethodNFEDeblurSRRandom inp.Box inp.
DiffPIR10034.83 / 0.93831.87 / 0.89232.45 / 0.92430.48 / 0.918
Flower5-OT50035.65 / 0.95433.03 / 0.93133.97 / 0.95231.85 / 0.952
SNaP, M = 10010035.91 / 0.95634.16 / 0.94634.84 / 0.95933.69 / 0.961

On fastMRI, 4 draws already exceed every posterior sampler in Table 3 in PSNR.

3. The MSE regressor. It is trained to minimise MSE, so it approximates the posterior mean, which is PSNR-optimal by construction, and a higher PSNR at few draws is expected. It is a point estimate with no diversity or uncertainty, so it is a reference for the best deterministic estimate, not a like-for-like competitor to a posterior sampler.

Proposition 2 holds for any non-degenerate Gaussian source, including the isotropic sources of Table 5, which nonetheless collapse in practice. So Proposition 2 does not support the specific source design. This should be emphasized in the text.

We add a discussion after Proposition 2 stating that it holds for any non-degenerate Gaussian source and therefore does not single out ours.

Proposition 2 guarantees that the one-step framework is correct; the choice of our source rests on Proposition 1, Lemma 2 and the empirical evidence.

The paper provides three complementary pieces of evidence for one-step posterior sampling, and Proposition 2 is not meant to be read alone.

  • Correctness (Proposition 2). The exact one-step map transports the designed source to the posterior, so the framework samples the right target.
  • Choice of source (Proposition 1 and Lemma 2, App. B.9). The mean of our source minimizes the expected transport displacement among linear estimators, and the source has a lower irreducible regression error than wider sources, including isotropic ones.
  • Practice (Sec. 4). SNaP is competitive with iterative samplers at a fraction of their inference compute. Measurement-adapted sources keep the draws diverse, while zero-mean isotropic sources collapse (Table 5).

SNaP is trained on paired data for each operator, yet is compared mainly with zero-shot methods, and the reported speedups exclude training. Compare with other per-operator one/few-steps posterior samplers, such as recent bridge models. They have similar advantages (single/few steps + diversity rather than point estimate).

Prompted by the reviewer's comment, we add a bridge model trained on the same pairs, I2SB, on all four CelebA tasks (100 test images).

With 4 draws, SNaP achieves higher PSNR and SSIM than I2SB on every task, using approximately 250× fewer NFEs.

I2SB runs at 999 NFE with a 105M-parameter network, against 9.1M for SNaP.

PSNR / SSIM / LPIPS.

MethodNFEDeblurSR ×2Random inp. (70%)Box inp.
I2SB99933.25 / 0.924 / 0.01331.92 / 0.913 / 0.01532.80 / 0.938 / 0.01128.72 / 0.804 / 0.043
SNaP, M = 1132.90 / 0.919 / 0.01831.27 / 0.903 / 0.02331.98 / 0.928 / 0.01830.73 / 0.934 / 0.021
SNaP, M = 4434.98 / 0.947 / 0.02233.27 / 0.935 / 0.02433.97 / 0.951 / 0.01632.78 / 0.954 / 0.019

I2SB has better LPIPS on three tasks, at 250× the evaluations. Per-operator training is paid once, while inference is paid on every reconstruction, which is where SNaP's speedup applies.

We have revised the paper to add this comparison to Table 1 and per-operator training to the Limitations.

Partial source ablation. Table 5 compares only zero-centered isotropic sources with the full source, so the contributions of the center mym_y and of the anisotropic covariance WW cannot be separated. Please complete the source ablation (Table 5) with N(my,τ2I)\mathcal{N}(m_y, \tau^2 I) and N(0,τ2W)\mathcal{N}(0, \tau^2 W).

Following the reviewer's suggestion, we state the choice of τ\tau in Sec. 3, move its sensitivity study to Sec. 4.3 (new Table 6), and report calibration on every dataset in Table 10 (App. B.4).

τ\tau affects reconstruction quality slightly but not calibration (Δcal\Delta_{\mathrm{cal}} 0.99–1.03). SNaP is calibrated on CelebA and AFHQ-Cat (0.88–1.08) and under-dispersed on fastMRI, which we state as a limitation.

1. Effect of τ\tau

τ\tauTransport energy (rel. to min.)PSNR, M=1M=1LPIPS, M=1M=1Δcal\Delta_{\mathrm{cal}} (→1)
0.15 (used)1.0632.940.0191.03
0.57.132.920.0190.99
1.023.932.510.0240.99

From τ=0.15\tau = 0.15 to 1.01.0, the transport energy grows from 1.06 to 23.9 times its minimum and PSNR drops by 0.43 dB, while Δcal\Delta_{\mathrm{cal}} stays within 0.99–1.03, so τ\tau need not be tuned precisely. Table 6 uses the measurements of the sweep in Table 14, so its τ=0.15\tau = 0.15 row differs slightly from Table 1.

2. Calibration on every dataset

DatasetCellsΔcal\Delta_{\mathrm{cal}} (→1)
CelebAdeblur / SR / random inp. / box inp.1.03 / 0.99 / 0.91 / 1.08
AFHQ-Catdeblur / SR / random inp. / box inp.1.02 / 0.96 / 0.88 / 1.08
fastMRI brain×4 20 dB / ×4 30 dB / ×8 20 dB / ×8 30 dB0.21 / 0.33 / 0.35 / 0.26

On fastMRI the draws are under-dispersed. We investigate this phenomenon in (App. B.4.) and Sec. 4.1 and Limitations.

Reviewer_5BSFOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SNaP is a MeanFlow-based posterior sampler for linear inverse problems with Gaussian noise. Its main contribution is a measurement-adapted Gaussian starting distribution, which limits variation in well-measured directions while preserving it in less constrained directions. Each sample requires one source solve and one network evaluation. Experiments evaluate reconstruction quality, diversity, and speed on natural-image restoration and multi-coil MRI, supplemented by synthetic problems with known posteriors.

Critical strengths

A useful and well-motivated source design. The starting distribution incorporates the measurement operator and noise level without requiring a separately trained source model. Table 5 shows that the complete design preserves sample diversity, whereas the tested isotropic sources produce nearly identical reconstructions. This is a meaningful improvement over simply conditioning MeanFlow on the observation.

Critical weaknesses
  1. The accuracy of the learned posterior remains a concern. The MRI results point the same way. If the draws were calibrated, averaging 16 draws should raise PSNR by about 2.75 dB over one draw, and CelebA matches this (2.78 dB on deblurring). On fastMRI the gain is only 0.79 to 1.24 dB (Table 3), which implies the draw variance is only about a quarter to a third of what it should be.
  2. The comparison mixes two settings. SNaP trains a separate model for each operator on paired data (App. C.3, 3.5 to 19 hours on 4 A6000 GPUs per model). All baselines in Tables 1 to 4 are zero-shot samplers that are never retrained. The only supervised baseline is an MSE regressor, which gives a point estimate. Supervised posterior samplers trained on the same pairs (conditional diffusion or bridge models) are not compared, and per-operator training is not listed as a limitation. This matters because the speed and quality claims rest on this comparison.
Questions for authors

1.What causes the variance inflation in Table 14 when the source is already the correct posterior? Does this error persist across training seeds or with longer training?

  1. Could you add the N(my,τ2I)\mathcal N(m_y,\tau^2I) baseline under the same architecture and training budget, reporting sample diversity and calibration alongside reconstruction quality?
6 comments

We thank the reviewer for feedback. In response, we have evaluated the calibration and spatial accuracy of SNaP's uncertainty estimates on MRI, alongside the baselines (W1). We have also compared SNaP with a bridge model trained on the same data pairs and with consistency-model solvers, and state per-operator training as a limitation (W2). We also rerun the single-Gaussian check over five seeds and with longer training (Q1), and complete the source ablation with both requested sources (Q2). See our point-by-point response below.

The accuracy of the learned posterior remains a concern. The MRI results point the same way. If the draws were calibrated, averaging 16 draws should raise PSNR by about 2.75 dB over one draw, and CelebA matches this (2.78 dB on deblurring). On fastMRI the gain is only 0.79 to 1.24 dB (Table 3), which implies the draw variance is only about a quarter to a third of what it should be.

We thank the reviewer for this observation, which motivated a closer examination of MRI, the only setting where SNaP exhibits miscalibration. Across all 8 CelebA and AFHQ-Cat tasks, SNaP achieves Δcal=0.88\Delta_{\mathrm{cal}}=0.88–1.081.08, whereas on MRI, Δcal=0.21\Delta_{\mathrm{cal}}=0.21–0.350.35, indicating under-dispersed posterior samples.

Our examination of MRI suggests that SNaP captures the spatial distribution of uncertainty but underestimates its magnitude. As shown in Fig. 5 of the revised paper, regions with larger reconstruction errors also exhibit greater sample variability. The correlation between per-pixel uncertainty and reconstruction error is comparable to that of the baselines. Moreover, rescaling the deviations of the samples from their mean by a single scalar, calibrated using the other test slices to achieve Δcal=1\Delta_{\mathrm{cal}}=1, brings SNaP's empirical coverage close to that of DPS, the only approximately calibrated baseline, without changing the mean reconstruction (4× MRI, 20 dB, 10 test slices, 16 samples per method).

DPSDAPSPnP-DMSNaP
Δcal\Delta_{\mathrm{cal}} (→1)1.020.500.790.22
Correlation of spread and error0.440.350.360.41
90% coverage, raw → one scalar88.4 → 88.173.8 → 87.384.0 → 88.056.4 → 87.6

We hypothesize that this under-dispersion arises from the multi-coil MRI measurement model, where noise is introduced in k-space, and how this noise is represented in SNaP's source distribution. We leave its correction to future work.

Thus, SNaP's MRI uncertainty estimates identify regions of higher reconstruction error, but underestimate the magnitude of uncertainty; a single global rescaling substantially improves calibration.

We have revised the paper to include this analysis in App. B.4 (Table 10 and 11, Fig. 5), and state the MRI under-dispersion in Sec. 4.1 and in the Limitations.

The comparison mixes two settings. SNaP trains a separate model for each operator on paired data (App. C.3, 3.5 to 19 hours on 4 A6000 GPUs per model). All baselines in Tables 1 to 4 are zero-shot samplers that are never retrained. The only supervised baseline is an MSE regressor, which gives a point estimate. Supervised posterior samplers trained on the same pairs (conditional diffusion or bridge models) are not compared, and per-operator training is not listed as a limitation. This matters because the speed and quality claims rest on this comparison.

Prompted by the reviewer's comment, we add a bridge model trained on the same pairs (I2SB) and few-step consistency-model solvers in two settings, and we now list per-operator training as a limitation.

With 4 draws, SNaP has higher PSNR and SSIM than I2SB on every task at 250× fewer evaluations, and at the same 4 NFE it outperforms PnP-CM in both settings and CM4IR on all but one metric.

1. Bridge model trained on the same pairs (I2SB). CelebA, 100 test images, PSNR / SSIM / LPIPS.

MethodNFEDeblurSR ×2Random inp. (70%)Box inp.
I2SB99933.25 / 0.924 / 0.01331.92 / 0.913 / 0.01532.80 / 0.938 / 0.01128.72 / 0.804 / 0.043
SNaP, M = 1132.90 / 0.919 / 0.01831.27 / 0.903 / 0.02331.98 / 0.928 / 0.01830.73 / 0.934 / 0.021
SNaP, M = 4434.98 / 0.947 / 0.02233.27 / 0.935 / 0.02433.97 / 0.951 / 0.01632.78 / 0.954 / 0.019

SNaP achieves higher PSNR and SSIM across all four tasks with approximately 250× fewer NFEs. I2SB achieves lower LPIPS on three tasks.

2. PnP-CM adapted to our setting. We distill consistency models on our CelebA and fastMRI training splits and run PnP-CM at 4 NFE with our operators. PSNR / LPIPS on CelebA, PSNR / SSIM on fastMRI.

TaskPnP-CM (4 NFE)SNaP, M=1 (1 NFE)SNaP, M=4 (4 NFE)
CelebA deblur33.60 / 0.04632.90 / 0.01834.98 / 0.022
CelebA SR ×229.26 / 0.06331.27 / 0.02333.27 / 0.024
CelebA random inp.29.00 / 0.06331.98 / 0.01833.97 / 0.016
CelebA box inp.27.82 / 0.05330.73 / 0.02132.78 / 0.019
MRI ×4, 20 dB30.50 / 0.73432.07 / 0.89032.68 / 0.905
MRI ×4, 30 dB30.60 / 0.72932.89 / 0.89933.80 / 0.920
MRI ×8, 20 dB25.13 / 0.58228.11 / 0.81929.07 / 0.844
MRI ×8, 30 dB25.14 / 0.58129.12 / 0.84529.89 / 0.865

At matched 4 NFEs, SNaP achieves higher PSNR and better perceptual or structural metrics in all eight settings. Even with a single NFE, SNaP achieves higher PSNR in seven of eight settings.

3. SNaP adapted to PnP-CM's setting. We train SNaP on PnP-CM's operators and use its released CelebA-HQ 256 checkpoint and 300 test images for PnP-CM and CM4IR (our rerun reproduces PnP-CM's paper within 0.15 dB). PSNR / LPIPS-VGG.

TaskPnP-CM (4 NFE)CM4IR (4 NFE)SNaP, M=1 (1 NFE)SNaP, M=4 (4 NFE)
SR ×4 (bicubic)27.27 / 0.28527.88 / 0.28925.41 / 0.32827.51 / 0.275
Deblur28.93 / 0.24928.30 / 0.29527.37 / 0.25429.41 / 0.217
Inpainting (70%)29.08 / 0.20429.55 / 0.28028.28 / 0.24230.38 / 0.198

At matched 4 NFEs, SNaP achieves the best performance on every metric except SR ×4 PSNR, where CM4IR is 0.37 dB higher.

Unlike methods that reuse a pretrained prior across operators, SNaP requires training for each forward operator. However, fixed forward operators are common in inverse problems, where the same acquisition setup is used to reconstruct many measurements, as in standardized MRI protocols. In such settings, the training cost is incurred once and amortized over repeated reconstructions, while each posterior draw requires at least 30× fewer NFEs than the posterior samplers in Table 4.

We have revised the paper to add these comparisons to App. B.2 and per-operator training to the Limitations.

What causes the variance inflation in Table 14 when the source is already the correct posterior? Does this error persist across training seeds or with longer training?

Prompted by the reviewer's comment, we reran the single-Gaussian check of Table 22 (Table 14 in the submission) over 5 training seeds at the paper's budget (20k steps). The table gives the error of the learned variance relative to the posterior variance.

Seed01234
Observed direction+50.4%−3.5%+3.1%−39.3%−2.5%
Null direction+52.6%+2.9%−1.1%−33.6%−1.6%

For three of the five seeds, the variance error is within 3.5% in both directions. The remaining two seeds exhibit substantial over- or underestimation, suggesting that the discrepancy is driven by training variability rather than a consistent directional bias.

Could you add the N(my,τ2I)\mathcal{N}(m_y, \tau^2 I) baseline under the same architecture and training budget, reporting sample diversity and calibration alongside reconstruction quality?

Prompted by the reviewer's comment, we add N(my,τ2I)\mathcal{N}(m_y, \tau^2 I), and also N(0,τ2W)\mathcal{N}(0, \tau^2 W) to complete every combination of mean and covariance, at the same architecture and training budget (CelebA deblurring, 100 test images). Alongside quality, we report diversity (spread), calibration and the transport energy E[∣∣x−x0∣∣2]\mathbb{E}[||x - x_0||^2].

Only the zero-mean isotropic sources collapse. Adding either the mean mym_y or the covariance WW keeps the draws diverse and calibrated, and the mean shortens the transport by an order of magnitude.

SourceEnergyPSNR M=1M=1PSNR M=100M=100LPIPS M=1M=1SpreadΔcal (→1)
N(0,I)\mathcal{N}(0, I)1.38135.8535.860.0310.25%0.00
N(0,τ2I)\mathcal{N}(0, \tau^2 I)0.40435.8435.850.0310.26%0.00
N(0,τ2W)\mathcal{N}(0, \tau^2 W)0.40032.9735.850.0185.42%0.96
N(my,τ2I)\mathcal{N}(m_y, \tau^2 I)0.03232.8935.920.0195.53%1.02
N(my,τ2W)\mathcal{N}(m_y, \tau^2 W) (ours)0.02832.9435.910.0195.50%1.03

The requested N(my,τ2I)\mathcal{N}(m_y, \tau^2 I) is diverse (spread 5.53%) and calibrated (Δcal 1.02). The full source achieves the lowest transport energy while maintaining comparable diversity, calibration, and reconstruction quality.

We have revised Table 5 and its discussion accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.