acceptodds
Under review as a conference paper at ICLR 2027

A2PS: Solving Noisy Inverse Problems via Asymptotically Accurate Posterior Sampling in Diffusion and Flow Models

Abstract

Training-free posterior sampling enables pretrained diffusion and flow models to solve inverse problems without task-specific retraining, but approximation errors can accumulate along the generative trajectory and bias the final distribution. We propose A2PS, a posterior sampling framework that progressively corrects these errors. Our key insight is that, near the end of generation, the distribution of the final clean sample conditioned on an intermediate state becomes increasingly well approximated by a Gaussian distribution, yielding an asymptotically accurate posterior score estimate. A2PS uses this estimate with Langevin dynamics to correct errors inherited from earlier steps. We establish the asymptotic accuracy of the posterior score and validate the resulting error-correction mechanism in a controlled setting. Across six noisy linear and nonlinear image inverse problems on ImageNet and FFHQ, with both pixel-space and latent-space models, A2PS achieves the best or near-best restoration quality among evaluated methods while also being the fastest across all tested tasks.

Primary area:
generative models
Keywords:
Diffusion Model, Flow Matching, Inverse Problem, Posterior Sampling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kSTiOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The authors propose a posterior sampling method for mitigating error accumulation in noisy diffusion trajectories. The key idea is to combine the posterior estimate towards the end of the diffusion sampling process with langevin dynamics to correct for the errors in the earlier steps. The authors show experimental results on a toy setup and on six inverse problems with linear and non-linear operators.

Critical strengths
  1. The toy experiments are quite intuitive and informative. The comparison between A2PS and A2PS (without re-eq) clearly establishes the error-correcting properties of the method (at least on the 2d toy dataset). I appreciate the authors' effort to demonstrate these properties intuitively for toy datasets, which is usually missing in a lot of related work.

  2. The method is training-free, scales to large diffusion backbones like SD-3.5 and has good quality compared to baselines (however, this is not fully convincing. See weaknesses below).

Critical weaknesses

Novelty and Related Work: The paper has certain claims that read as its own contribution but have already been explored in related work. This casts doubt on the original contributions of the paper. For instance,

– To the best of my knowledge, the ideas in Section 3.1 and 3.2 of the paper are not new. These results have already been established in TMPD [1]. Eq. 25 in Lemma D.1 in this paper is the same as Proposition 1 in the TMPD paper. However, the TMPD paper is not cited and the several theorems and lemmas stated in this paper make it seem like these contributions are novel.

– The idea behind Eq. 9 proposed in Proposition 3.2 is not new. Optimizing the measurement error subject to manifold constraints like maintaining proximity between the guided and the unguided scores has been proposed in a number of works like DiffPIR [2] and NDTM [3]. However, NDTM and related methods are not cited.

– Can the authors clarify more about what their original contributions are, since an SGD-like optimization update is common in methods like NDTM and RB-Modulation [4]

Empirical Comparisons: I think the empirical comparisons are dubious.

– Table 1 lacks some important baselines like NDTM and RB-Modulation. Can the authors present additional comparisons with these baselines?

– Some of the runtime costs for the baselines are inflated. For instance, the appendix mentions DDRM run with 500 diffusion steps, but the official implementation uses 20 steps. This largely inflates the cost of DDRM. Can the authors clarify more on this?

– Discussion of reconstruction metrics like PSNR and SSIM is missing. It is important to discuss the perception-reconstruction tradeoff for different methods and not just the perceptual quality.

Questions for authors

See weaknesses

6 comments

Clarification on Novelty and Related Work

We thank the reviewer for pointing out the similarity with prior work. We would like to clarify the key distinctions between these works and ours, and further elaborate on the novel contributions of our approach.

Gaussian approximations of the clean conditional via Tweedie's formula (Eq.7 and 9) have been studied in several works [1 - 4, 9]. In particular, [1 - 4] utilize the same posterior approximation *(1):p(x1∣xt=x)≈N(x1;μ1∣t,Σ1∣t)p(\mathbf{x}_1 \mid \mathbf{x}_t = \mathbf{x}) \approx \mathcal{N}(\mathbf{x}_1;\boldsymbol{\mu}_{1\mid t},\boldsymbol{\Sigma}_{1\mid t}) to estimate the ideal posterior score st∣ys_{t|y}. Also, this approximation naturally leads to the similar optimizing target to [2 - 4] that maintain the proximity between the guided and the unguided scores.

However, without further analyze the discrepancy induced by this approximation, we cannot guarantee accurate posterior sampling. That is, even if the generating process perfectly follows an approximate score, the gap to the real st∣ys_{t|y} can lead to a biased target distribution.

This motivates our asymptotic analysis in Sec. 3.2, which establishes the following asymptotic chain as t→1t\to1: \left. \begin{aligned} p(\boldsymbol{x}_1 \mid \mathbf{x}_t = \boldsymbol{x}) &= \mathcal{N}(\boldsymbol{x}_1; \boldsymbol{\mu}_{1|t}, \boldsymbol{\Sigma}_{1|t}) \left[1 + \mathcal{O}(\rho_t^3)\right] \\\\ \tilde{\boldsymbol{\mu}}_{1|t,y} &= \boldsymbol{\mu}_{1|t,y} + \mathcal{O}(\rho_t^4) \end{aligned} \right\\} \Longrightarrow \tilde{\boldsymbol{s}}_{t|y} = \boldsymbol{s}_{t|y} + \mathcal{O}(\rho_t^2)
Significantly, this asymptotic chain provide a surrogate posterior score s~t∣y\tilde{\boldsymbol{s}}_{t|y} that effectively minimize the gap to the target distribution in the end by Langevin Dynamics (Sec. 3.3).

In summary, our asymptotic analysis establishes that our approximate score s~t∣y\tilde{s}_{t\mid y} asymptotically minimizes the discrepancy between the sampled distribution and the target distribution, providing a theoretical guarantee absent from the aforementioned approximations.

We acknowledge the previously mentioned similarity and have revised the paper accordingly to better clarify the relationship to these works.

*(1) TMPD ([1]) specifically derive the exact covariance:Σ1∣t(x)=σ02at2bt∂μ1∣t(x)∂x.\boldsymbol{\Sigma}_{1\mid t}(\mathbf{x}) = \frac{\sigma_0^2 a_t^2}{b_t}\frac{\partial\boldsymbol{\mu}_{1\mid t}(\mathbf{x})}{\partial\mathbf{x}}. Our original submission already cited [8] in the discussion of Tweedie's formula, which also establishes the corresponding covariance identity in the one-dimensional setting. Since our work considers the multi-dimensional setting, we have additionally cited [10] to acknowledge the more directly relevant derivation that provide TMPD's result.

Empirical Comparisons to Other Baselines

We thank the reviewer for emphasizing the importance of these additional comparisons. In response, we conducted experiments with NDTM [3], RB-Modulation [4], and DiffPIR [2] across all six inverse problems considered in Table 1, using 1,000 ImageNet validation images per task. For all three baselines, we followed the hyperparameter settings recommended in their respective papers or official implementations, without additional task-specific tuning. The results are summarized below.

Additional quantitative comparisons (LPIPS ↓ / FID ↓ / runtime ↓ in seconds per image):

MethodBox InpaintingCheckerboardRandom 70%Super-resolutionGaussian DeblurringNonlinear Deblurring
DiffPIR.255/90.42/ 0.56.119 / 56.20 / 0.57.119 / 36.19 / 0.56.243 / 37.09 / 2.83.360 / 77.26 / 2.88*(1)
NDTM.251 / 79.77 / 14.81.278 / 59.41 / 14.87.063 / 16.29 / 14.81.246 / 48.84 / 14.75*(1).155 / 32.28 / 14.91
RB-Modulation.297 / 96.42 / 14.83.354 / 102.00 / 14.85.126 / 26.98 / 14.77.206 / 39.13 / 14.79*(1).223 / 43.20 / 14.90
A2PS (Ours)*(2).189 / 53.00 / 1.72.113 / 23.51 / 1.71.119 / 31.94 / 1.71.224 / 41.89 / 1.71.243 / 40.75 / 2.75.169 / 39.35 / 4.62
A2PS (Ours).163 / 45.35 / 7.36.088 / 20.00 / 7.36.074 / 23.28 / 7.36.184 / 39.63 / 7.42.222 / 40.56 / 15.02.157 / 37.54 / 39.78

The additional results provide shows that our method achieves the best LPIPS on four out of six tasks and the best FID on three out of six tasks among the four methods.

Overall, these additional experiments strengthen our empirical evaluation against the suggested baselines and further demonstrate the effectiveness of A2PS across diverse inverse problems.

*(1) We didn't find the corresponding parapeters for these method on these tasks.

*(2) One step Langevin Dynamic

Discussion of resconstruction metrics like PSNR and SSIM

We thank the reviewer for raising this important point. We agree that PSNR and SSIM provide valuable measures of reconstruction fidelity and that their omission leaves an incomplete picture of the perception–distortion trade-off.

However, since inverse problem is generally ill-posed, pixel-level fidelity may not fully reflect perceptual quality ([5, 7]). Moreover, these metrics may prefer some plausible image details rather than faithfully reflect the image quality ([6]). Therefore, we chose LPIPS and FID to be our main metrics. Detail of this discussion is provided in the revised scripts Appendix C, including figures that showcase examples when high PSNR and SSIM can exhibit lower perceptual quality in certain cases.

To make the evaluation complete, here we report the PSNR and SSIM results on ImageNet below.

PSNR ↑ (dB) / SSIM ↑

MethodBox InpaintingCheckerboard InpaintingRandom 70% InpaintingSuper ResolutionGaussian DeblurringNon-linear Deblurring
piGDM16.96 / .64121.98 / .60723.56 / .58723.18 / .548*(1)*(1)
DiffPIR19.31 / .66823.82 / .69325.51 / .69324.30 / .62722.30 / .421*(1)
NDTM15.47 / .72417.81 / .63726.19 / .77321.99 / .467*(1)24.41 / .657
RB-Modulation16.76 / .62517.34 / .51825.40 / .72322.81 / .521*(1)22.85 / .579
A2PS (Ours)19.82 / .72223.40 / .70125.11 / .70022.65 / .50923.43 / .54924.28 / .573

Although A2PS does not consistently achieve the highest PSNR or SSIM, its reconstruction fidelity remains competitive with existing methods. Moreover, we've included additional qualitative results in the revised manuscripts Appendix G, showing the good quality of our generation.

*(1) We didn't find the corresponding parapeters for these method on these tasks.

Runtime Fairness and Baseline Sampling Steps

We thank the reviewer for raising this concern. We used longer sampling schedules (500 steps for DDRM and 250 steps for DDNM+) to evaluate baseline performance beyond their fast configurations, rather than to artificially inflate their runtimes. Nevertheless, we acknowledge that reporting only these configurations provided an incomplete quality–runtime comparison.

To address this, we conducted additional experiments with DDRM (20 steps) and DDNM+ (100 steps). We also evaluated A2PS with only one Langevin dynamics (LD) step, alongside our original five-step configuration, to examine the quality–runtime trade-off.

Additional Results (LPIPS ↓ / FID ↓ / Runtime per image ↓, s)

MethodNFEsBox InpaintingCheckerboard InpaintingRandom 70% InpaintingSuper ResolutionGaussian Deblurring
DDRM (20 steps)20.230 / 69.90 / 1.25.154 / 38.77 / 1.38.254 / 76.02 / 1.18.264 / 42.80 / 0.78.382 / 62.47 / 0.80
DDNM+ (100 steps)100.197 / 54.10 / 3.45.091 / 21.12 / 3.58.106 / 30.72 / 3.39.251 / 43.25 / 2.96.776 / 358.1 / 3.11
A2PS(Ours)*(1)40.189 / 53.00 / 1.72.113 / 23.51 / 1.71.119 / 31.94 / 1.71.224 / 41.89 / 1.71.243 / 40.75 / 2.75
A2PS(Ours)120.163 / 45.35 / 7.36.088 / 20.00 / 7.36.074 / 23.28 / 7.36.184 / 39.63 / 7.42.222 / 40.56 / 15.02

*(1) One step Langevin Dynamic

Notably, A2PS(1 step LD) achieves lower LPIPS and FID than DDRM (20 steps) across all five tasks, while maintaining seconds-scale inference. Increasing LD steps to 5 further improves reconstruction quality at additional computational cost.

Although A2PS is not uniformly faster than the fastest DDRM configuration, these results demonstrate its favorable quality–runtime trade-off, even with minimal Langevin correction. We will incorporate these comparisons, clarify our sampling-step choices, and revise our efficiency claims accordingly.

Reference

[1] Boys, Benjamin, et al. "Tweedie moment projected diffusions for inverse problems." arXiv preprint arXiv:2310.06721 (2023).

[2] Zhu, Yuanzhi, et al. "Denoising diffusion models for plug-and-play image restoration." 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). IEEE, 2023.

[3] Pandey, et al. "Variational Control for Guidance in Diffusion Models." Proceedings of the 42nd International Conference on Machine Learning. PMLR, 2025.

[4] Rout, et al. "RB-Modulation: Training-Free Stylization using Reference-Based Modulation." The Thirteenth International Conference on Learning Representations. 2025.

[5] Christian Ledig, et al. Photo-realistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 105–114. IEEE, 2017.

[6] Zhang, et al. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 586–595. IEEE, 2018.

[7] Yochai Blau and Tomer Michaeli. The Perception-Distortion Tradeoff. pp. 6228–6237, 2018.

[8] Bradley Efron. Tweedie’s formula and selection bias. Journal of the American Statistical Association, 106(496):1602–1614, 2011.

[9] Chung, et al. "Diffusion Posterior Sampling for General Noisy Inverse Problems." The Eleventh International Conference on Learning Representations . 2023.

[10] Chenlin Meng, et al. "Estimating high order gradients of the data distribution by denoising." In Proceedings of the 35th International Conference on Neural Information Processing Systems (NeurIPS). 2021.

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

The manuscript proposes A2PSA^2PS, a new technique for correcting approximation errors along the generative trajectory of a flow/diffusion model used to solve an inverse problem. The technique works by applying Langevin dynamics together with an asymptotically accurate characterization of the posterior that becomes more accurate towards the end of the generation process. The principal use case for this technique is enabling more accurate posterior sampling for noisy inverse problems. The work is validated experimentally by a 2D toy model study and experiments on ImageNet and FFHQ with six degradation models and flow and diffusion backbones.

Critical strengths

The method is conceptually well-motivated and well-designed. The manuscript clearly presents what appears to be an effective practical solution to the issue of compounding trajectory errors in posterior sampling for inverse problem solving.

Critical weaknesses
  1. Firstly, the claim that the method empirically improves posterior sampling for real-world problems is only weakly supported. The in-depth 2D validation is only performed with a single example posterior, which may limit the results’ repeatability. For the image results, the FID and LPIPS metrics don’t directly measure posterior quality. Additionally, without representative samples from each of the methods, it is difficult to interpret the FID and LPIPS scores and to be convinced that the method has indeed improved posterior sampling.

  2. Secondly, the experimental setup leaves questions about the fairness relative to other approaches. For example, it is claimed that A^2PS “is the fastest method across all evaluated tasks”; however, competing methods are evaluated with more neural function evaluations (NFEs), and the number of NFEs only appears to have been optimized for A^2PS. Similarly, it is not clear that the hyperparameters of baseline methods have been optimized for the tasks measured in this work (and the LPIPS and FID metrics).

Questions for authors
  1. [Future work] LPIPS and FID are useful metrics for measuring perceptual fidelity to ground truth and the marginal realism of generated images (respectively), but neither metric concretely assesses posterior accuracy. Given that the focus of the paper is accurate posterior sampling, could you include metrics such as calibration (i.e., do x% credible intervals contain the ground truth exactly x% of the time?) or proper scoring rules?
  2. [Future work] Please include a visualization of example repeated samples (with different noisy seeds but the same inverse problem task and measurement noise) from each of the methods for some of the noisy imaging inverse problems, so that the relative image quality can be more easily assessed.
  3. [Future work] Do the results of the 2D posterior hold true with different choices of prior and measurement guidance?
  4. [Clarification] Please clarify the computational cost of your algorithm, relative to others. For example, how long do the Langevin steps take (and how does this depend on the cost of the measurement operator)? How long does the initial 10,000-sample estimation of the backbone’s noise parameters take?
  5. [Clarification / Future work] The motivation for this work is reducing errors inherited from earlier sampling steps. While the mathematics supports this explanation, it would be instructive to see this effect empirically, e.g., measuring the improvement in trajectory when starting from an incorrect noisy point.
  6. [Clarification] Many papers attempt to start diffusion processes partway through their trajectory (e.g., with approaches like starting a reverse diffusion process partway along via a noised version of a cheap image estimate), where it is unclear that such a starting point is actually on the diffusion trajectory. Could your re-equilibration strategy also find applicability to this sister problem?
10 comments

Response to W1 and Q1–3: Further Evidence Regarding Improvement in Posterior Sampling

We thank the reviewer for highlighting the importance of directly evaluating posterior sampling quality beyond perceptual reconstruction metrics, and for suggesting additional experiments to assess the robustness and diversity of the generated samples. To address these concerns, we have conducted additional controlled 2D experiments with diverse priors and measurement operators, included repeated-sample visualizations for image restoration tasks, and introduced an additional quantitative evaluation using the Energy Score, a proper scoring rule for probabilistic forecasts ([1]) and posterior coverage. The additional experiments all shows that A2PS consistently surpass baselines.

1. Additional 2D Experiments with Diverse Priors and Measurement Operators (Q3)

To examine whether the observations in the original 2D experiment generalize beyond the specific prior and linear measurement setting, we have conducted additional experiments with diverse prior distributions and measurement operators, now included in revised script's Appendix A.2 with implementation details and density visualizations. Specifically, we consider imbalanced multimodal Gaussian-mixture and two-armed spiral priors, together with quadratic and nonlinear saturating measurement operators based on the hyperbolic tangent function.

We evaluate all combinations of the considered priors and measurement operators, reporting the terminal total variation (TV) distance and forward KL divergence with respect to the corresponding reference posterior distributions. Across the additional configurations, A2PS consistently reduces the posterior discrepancies to the numerical integration floor.

These results provide further empirical evidence for the robustness of A2PS across the different prior distributions and measurement operators.

2. Repeated-Sample Visualizations for Image Restoration (Q2)

We agree that repeated samples conditioned on the same observation provide useful qualitative evidence of sample variability and help assess whether a method produces diverse reconstructions or tends to concentrate on a limited set of outputs.

Following the reviewer's suggestion, we have added representative repeated-sample visualizations in Appendix G. For each inverse problem, we generate multiple samples conditioned on the same noisy observation, using different random seeds. These visualizations allow direct comparison of the reconstruction quality and sample diversity across methods, while illustrating whether the methods exhibit potential mode collapse.

The repeated-sample visualizations provide complementary qualitative evidence regarding the behavior of A2PS on image restoration problems. In particular, they illustrate how the methods represent variability across different random seeds, rather than evaluating each method solely through a single reconstruction.

3. Additional Posterior Evaluation Using a Proper Scoring Rule (Q1)

To further assess the statistical quality of generated posterior samples beyond perceptual metrics, we conduct additional quantitative experiments on three representative image restoration tasks: Checkerboard Inpainting and Super-resolution, involving linear operators, and Nonlinear Deblurring, involving a nonlinear operator.

Our evaluation uses 50 ImageNet-256 reference images for each task. For each reference image, we generate two independently noise-corrupted observations and obtain eight posterior samples per observation from each method. We compare A2PS against DPS and DAPS using the Energy Score. This results in 800 generated samples per method for each task, or 2,400 samples per method across all tasks.

Energy Score (ES). We employ the Energy Score, a multivariate proper scoring rule, to evaluate the quality of the generated sample ensembles. Unlike metrics based on a single reconstruction, the Energy Score accounts for both the discrepancy between generated samples and the reference image and the dispersion of the predictive distribution. In particular, it evaluates the generated ensemble as a probabilistic prediction.

For a set of MM generated samples x1,…,xMx_1,\ldots,x_M and ground truth x∗x^\ast, the empirical score is computed as ES^=1M∑i=1M∥∣x(i)−x∗∥∣2D−12M(M−1)∑i≠j∥∣x(i)−x(j)∥∣2D\widehat{\mathrm{ES}}=\frac{1}{M}\sum_{i=1}^{M}\frac{\||\mathbf{x}^{(i)}-\mathbf{x}^*\||_2}{\sqrt{D}}-\frac{1}{2M(M-1)}\sum_{i\neq j}\frac{\||\mathbf{x}^{(i)}-\mathbf{x}^{(j)}\||_2}{\sqrt{D}} where DD is the dimensionality of the flattened image. The normalization by D\sqrt{D} rescales the Euclidean distances and does not change the relative comparison between methods.

(cont'd)

The Energy Score is computed for each observation using its corresponding reference image and the generated sample ensemble.

As reported in Table 1, A2PS achieves the lowest Energy Score among the evaluated methods, outperforming both DPS and DAPS.

Table 1. Energy Score (lower is better).

TaskA2PS(Ours)DAPSDPS
Super-resolution (×4).047.051.090
Nonlinear Deblurring.035.039.071
Checkerboard Inpainting *(1).054.056.073

To quantify the image-wise consistency of the improvement over baselines, we compute paired Energy Score differences for each reference image, showing in Table 2. A2PS achieves lower scores than baselines for most of the images in all of the tasks.

Table 2. Paired Energy Score Difference Between A2PS and Baselines (A2PS − Baseline; lower is better)

TaskComparisonES Difference [95% CI]# of Reference Images Favoring A2PS
Super-resolutionDAPS−0.0055 [−0.0069, −0.0042]48/50
DPS−0.0413 [−0.0459, −0.0369]50/50
Nonlinear DeblurringDAPS−0.0061 [−0.0075,−0.0049]49/50
DPS−0.0345 [−0.0393, −0.0299]50/50
Checkerboard Inpainting *(1)DAPS−0.0046 [−0.0055, −0.0037]50/50
DPS−0.0206 [−0.0235, −0.0179]50/50

Posterior Coverage. We evaluate posterior uncertainty calibration using pixel-wise marginal coverage, which measures the fraction of ground-truth pixel values falling within the range of the generated posterior samples. For each pixel dd, the interval is defined as Id=[min⁡1≤i≤Mxd(i),max⁡1≤i≤Mxd(i)].I_d=\left[\min_{1\leq i\leq M} x_d^{(i)},\max_{1\leq i\leq M}x_d^{(i)}\right]. The empirical coverage is computed as C^=1D∑d=1DI[xd∗∈Id],\widehat{C} = \frac{1}{D}\sum_{d=1}^{D}\mathbb{I}[x_d^*\in I_d], where DD is the number of evaluated pixels. Under exchangeability and continuity, the expected coverage is (M−1)/(M+1)(M-1)/(M+1), yielding 7/9≈0.7787/9\approx 0.778 for M=8M=8 samples. Coverage substantially below this reference indicates underestimation of posterior uncertainty. We compute coverage per observation and aggregate across observations, accounting for the shared ground-truth image of the two observations.

Table 3. Pixel-wise Marginal Coverage of the Min–Max Interval with mean ±\pm std

Task/MethodA2PSDAPSDPS
Super-resolution.770 ±\pm .062.602 ±\pm .101.609 ±\pm .101
Nonlinear deblurring.802 ±\pm .042.439 ±\pm .052.634 ±\pm .091
Checkerboard *(1).781 ±\pm .020.706 ±\pm .087.559 ±\pm .089

A2PS obtains coverage values of 0.770, 0.802, and 0.781 across the three tasks, all close to the 0.778 reference. DPS and DAPS show lower coverage in these experiments, indicating that their sample ranges more frequently exclude the ground-truth pixel values.

Together, the Energy Score and coverage results provide complementary evidence that A2PS better captures the distributional properties and pixel-wise marginal uncertainty of posterior samples across these three image restoration tasks.

*(1) For Checkerboard Inpainting, coverage is evaluated only on unobserved pixels, i.e., pixels masked out from the input observation and requiring reconstruction.


Response to W2 and Q4: Fairness of NFE, Hyperparameter Settings, and Computational Cost

We thank the reviewer for raising these concerns. Our intention was to evaluate each baseline under its recommended or task-adapted configuration. Specifically, we followed the original authors' recommended hyperparameters where available and performed task-specific searches otherwise (Appendix C.1 in original script, Appendix D.1 in revised script).

Regarding computational fairness, we would like to argure that equal NFEs do not necessarily imply equal computational cost, as different methods involve different numbers of gradient evaluations, optimization steps, and measurement-operator evaluations. We therefore report end-to-end runtimes measured on the same GPU.

To provide a more comprehensive comparison, we additionally evaluated DPS, DAPS, and A2PS with different Langevin dynamics (LD) budgets (1 and 5 steps), keeping the outer generative schedule fixed at 20 steps.

Additional Results: LPIPS ↓ / FID ↓ / Runtime per image ↓ (s)

MethodNFEBox InpaintingCheckerboard InpaintingRandom 70% InpaintingSuper ResolutionGaussian DeblurringNonlinear-Deblurring
DPS1000.415 / 78.91 / 62.98.387 / 68.84 / 62.98.388 / 68.49 / 62.98.493 / 85.52 / 62.96.365 / 64.96 / 63.16.393 / 70.87 / 65.96
DAPS250.214 / 68.44 / 16.62.098 / 32.14 / 16.62.073 / 23.70 / 16.62.307 / 49.34 / 18.81.307 / 50.02 / 26.33.187 / 44.74 / 115.18
A2PS (1 step LD)40.189 / 53.00 / 1.72.113 / 23.51 / 1.71.119 / 31.94 / 1.71.224 / 41.89 / 1.71.243 / 40.75 / 2.750.169 / 39.35 / 4.62
A2PS (5 step LD)120.163 / 45.35 / 7.36.088 / 20.00 / 7.36.074 / 23.28 / 7.36.184 / 39.63 / 7.42.222 / 40.56 / 15.02.157 / 37.54 / 39.78

(cont'd)

Regarding Other Computational Cost

Computational cost of Langevin re-equilibration. Each of re-equilibration step include one NFE and respecting endpoint optimization, thus are task dependent. For example, the higher costs for Gaussian and nonlinear deblurring reflect the greater computational demands of their measurement operators and associated gradient evaluations. Nevertheless, using fast LD methods allows less LD steps, significantly reducing the computational cost.

Offline preprocessing cost. Regarding offline preprocessing, A2PS estimates the backbone-specific uncertainty profile from 10,000 prior trajectories using 20 sampling steps. This takes 1h 33m 24s for the pixel-space diffusion model and 3h 40m 36s for SD3.5-M. Importantly, this is a one-time computation for each pretrained backbone, reused across different inverse problems and observations, rather than an additional cost incurred for every reconstruction.


Response to Q5: showcase error correction mechanism empirically

We thank the reviewer for this constructive suggestion. We believe that our existing Fig. 4 already provides direct empirical evidence for the proposed error-reduction mechanism at the trajectory level.

Specifically, Fig. 4 tracks the TV and KL divergences between the sampling distribution and the oracle posterior throughout the entire sampling trajectory. The results reveal a clear distinction between methods with and without error-reduction mechanisms: for DPS (orange lines) and A2PS without re-equilibration (purple lines), the distributional errors continuously accumulate toward the end of sampling. In contrast, for DAPS (blue lines) and A2PS (green lines), the errors first increase due to deviations introduced in earlier sampling steps, but subsequently decrease as the error-reduction mechanism takes effect. This directly demonstrates that the proposed mechanism can correct distributional errors inherited from earlier steps, rather than merely improving the final reconstruction quality.

More importantly, A2PS exhibits substantially stronger error reduction than DAPS, bringing both TV and KL divergences close to the numerical floor. This behavior remains evident even under practical approximations, including finite Langevin steps (green dotted lines) and isotropic covariance approximation (pink lines), demonstrating that effective error correction does not rely on idealized full equilibration or exact covariance estimation.

Furthermore, in the revised manuscript, we have included additional experiments with different priors and measurement settings, which exhibit consistent error-reduction behavior. These results further demonstrate that the observed phenomenon is not specific to the original 2D experimental configuration.


Response to Q6: if the re-equilibrium mechanism can be applied to other diffusion problems

Yes. Our re-equilibration strategy can be directly applied to this setting. The key idea is to use Langevin dynamics to move an arbitrary initialization toward a desired target distribution, provided that its score is available.

For the setting described by the reviewer, the target distribution can be chosen as the pretrained diffusion model's marginal distribution pt(xt)p_t(x_t) at the selected intermediate time tt. Specifically, its score, ∇xtlog⁡pt(xt)\nabla_{x_t}\log p_t(x_t) , can be used to guide Langevin dynamics. Starting from a noised version of a cheap image estimate, the samples can be re-equilibrated toward the model's marginal distribution at that time before continuing the reverse diffusion process. This allows the initialization to be brought closer to the diffusion model's trajectory, reducing the distributional mismatch caused by an arbitrary starting point.

Furthermore, as demonstrated in our paper, fast Langevin dynamics provides an effective way to achieve re-equilibration with a limited computational budget, making this strategy practical for such intermediate-time initialization scenarios.

References

[1] Gneiting, Tilmann, and Adrian E. Raftery. "Strictly proper scoring rules, prediction, and estimation." Journal of the American statistical Association 102.477 (2007): 359-378.

Reviewer_Gc11OpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes A2PS, a training-free posterior-sampling framework for solving noisy inverse problems with pretrained diffusion and flow-matching models. The method derives an asymptotic Gaussian approximation to the clean endpoint conditional distribution, uses it to estimate the posterior score, and applies Langevin correction to correct errors accumulated along the generative trajectory. The paper provides theoretical analysis and evaluates the method on controlled and large-scale image-restoration problems.

Critical strengths
  1. The paper identifies an important problem in training-free inverse-problem solving: errors in approximate posterior guidance can accumulate over the generative trajectory.
  2. The asymptotic characterization of the endpoint conditional is mathematically interesting and provides a principled motivation for posterior-score estimation near the generation endpoint.
Critical weaknesses
  1. My immediate concern was regarding this work's novelty relative to Pi-GDM [1]. Pi-GDM already uses a Gaussian approximation to the clean conditional to derive measurement guidance. A2PS offers an asymptotic characterization and introduces Langevin re-equilibration, but the manuscript does not adequately differentiate the contribution of these components relative to existing guidance methods. In particular, the absence of direct comparisons with Pi-GDM makes it difficult to assess the incremental contribution. Moreover, this paper is not cited either, in spite of being a well-known work in this area.

  2. The additional likelihood-guidance rule in Eq. (12) has the same score-extrapolation structure as classifier-free guidance (CFG). I believe this connection should be acknowledged and its role distinguished from standard CFG.

  3. The theoretical result is local and asymptotic, whereas the implementation uses an offline isotropic covariance estimate. The manuscript should clarify the approximation error under this substitution. That is, is the method truly asymptotically accurate?

  4. Critical question: Although the paper reports quantitative restoration results across multiple degradation operators, model backbones, and datasets, it does not provide qualitative comparisons of actual restored images in the main results section. This omission makes it difficult to assess the visual quality of the reconstructions between A2PS and competing methods. I strongly recommend including representative side-by-side comparisons of the degraded inputs, ground-truth images, and outputs from A2PS and strong baselines across several inverse problems. Such comparisons are absolutely necessary and would complement the reported LPIPS and FID metrics and strengthen the empirical evidence for the method's effectiveness.

  5. Could the experiment shown in Fig. 2 be elaborated more? How was this simulation run and what is the density being considered? Kindly add these details at least to the appendix.

References:

  1. Song, Jiaming, Arash Vahdat, Morteza Mardani, and Jan Kautz. "Pseudoinverse-guided diffusion models for inverse problems." In International conference on learning representations. 2023.
Questions for authors

Please see weaknesses.

4 comments

Response to W1: Relation to Pi-GDM([1])

We thank the reviewer for pointing out the similarity with prior work. We would like to clarify the key distinctions between these works and ours, and further elaborate on the novel contributions of our approach.

Gaussian approximations of the clean conditional via Tweedie's formula (Eq.7 and 9) have been studied in several works [1–3]. In particular, Pi-GDM [1] and TMPD [2] utilize the same posterior approximation: p(x1∣xt=x)≈N(x1;μ1∣t,Σ1∣t)p(\mathbf{x}_1 \mid \mathbf{x}_t = \mathbf{x}) \approx\mathcal{N}(\mathbf{x}_1;\boldsymbol{\mu}_{1\mid t},\boldsymbol{\Sigma}_{1\mid t}) to estimate the ideal posterior score st∣ys_{t|y}.

However, without further analyze the discrepancy induced by this approximation, we cannot guarantee accurate posterior sampling. That is, even if the generating process perfectly follows an approximate score, the gap to the real st∣ys_{t|y} can lead to a biased target distribution.

This motivates our asymptotic analysis in Sec. 3.2, which establishes the following asymptotic chain as t→1t\to1: \left. \begin{aligned} p(\boldsymbol{x}_1 \mid \mathbf{x}_t = \boldsymbol{x}) &= \mathcal{N}(\boldsymbol{x}_1; \boldsymbol{\mu}_{1|t},\boldsymbol{\Sigma}_{1|t}) \left[1 + \mathcal{O}(\rho_t^3)\right] \\\\ \tilde{\boldsymbol{\mu}}_{1|t,y}&= \boldsymbol{\mu}_{1|t,y} + \mathcal{O}(\rho_t^4)\end{aligned}\right\\}\Longrightarrow\tilde{\boldsymbol{s}}_{t|y}=\boldsymbol{s}_{t|y} + \mathcal{O}(\rho_t^2) Significantly, this asymptotic chain provide a surrogate posterior score s~t∣y\tilde{\boldsymbol{s}}_{t|y} that effectively minimize the gap to the target distribution in the end by Langevin Dynamics (Sec. 3.3).

In summary, our asymptotic analysis establishes that our approximate score s~t∣y\tilde{s}_{t\mid y} asymptotically minimizes the discrepancy between the sampled distribution and the target distribution, providing a theoretical guarantee absent from the aforementioned approximations.

We acknowledge the previously mentioned similarity and have revised the paper accordingly to better clarify the relationship to these works.

Quantitative Comparison to Pi-GDM

We additionally evaluated Pi-GDM on the 4 inverse-problem tasks considered in our experiments, using the parameter configurations reported in the original Pi-GDM paper and the same experimental settings to compare with ours metrics.

TaskMethodLPIPS ↓FID ↓Runtime (s)
Box inpaintingPi-GDM0.23575.917.20
A2PS0.16345.357.36
Checkerboard inpaintingPi-GDM0.13836.027.19
A2PS0.08820.007.36
Random 70% inpaintingPi-GDM0.11535.347.19
A2PS0.07423.287.36
Super-resolutionPi-GDM0.29049.827.21
A2PS0.18439.637.42

A2PS consistently achieves better LPIPS and FID across all 4 tasks, demonstrating substantial improvements in both perceptual similarity and distributional quality.


Response to W2: Connection to Classifier-Free Guidance (CFG) of Eq. 9

We thank the reviewer for pointing out this connection. We initially viewed Eq. (12) as serving a different purpose from standard CFG, which motivated the omission of the citation. Standard CFG amplifies conditional guidance during generation, requiring additional training to learn condition-dependent guidance, while our Eq. (12) specifically reweights the measurement likelihood to strengthen data consistency in inverse problems and is training-free.

However, upon reconsideration, we agree that Eq. (12) shares the same score-extrapolation structure as CFG and that this connection deserves explicit acknowledgment. We have therefore added the CFG citation and clarified both the mathematical connection and the difference in their intended roles in the revised manuscript.


Response to W3:

We thank the reviewer for pointing out the distinction between our theoretical analysis and practical implementation. Since evaluating the full covariance matrix usually requires huge computational cost, assuming the isotropy of the covariance matrix and estimate it offline from prior trajectories is considered to be a more pratical way.

While our theoretical derivation does not assume an isotropic covariance, it retains the full state-dependent covariance and establishes that Σ1∣t(x)=ρt2I+O(ρt4),ρt→0 as t→1.\Sigma_{1|t}(x)=\rho_t^2 I+O(\rho_t^4), \qquad \rho_t\to0 \text{ as }t\to1. This result shows that the full covariance becomes asymptotically isotropic near the endpoint, providing a theoretical motivation for the isotropic covariance approximation adopted in Section 3.3.

The 2D experiments in Figure 4 (pink curve) demonstrate that the isotropic approximation can still yield an accurate posterior distribution while the prior remains anisotropic. Furthermore, this observation remain consistent with different prior and measurement settings, demonstrated in the newly provided 2D results in our revised script.

Response to W4: Qualitative Results

We thank the reviewer for this suggestion. Due to space constraints, qualitative comparisons were not included in the original submission. We have now substantially expanded the qualitative results in the revised manuscript Appendix G, including comparisons across different inverse problems and baselines. We kindly refer the reviewer to the revised paper for these additional visual comparisons and analyses.


Response to W5: experiment detail on 2D experiment (Fig. 2)

We thank the reviewer for pointing out the need for a more detailed description of the 2D experiment. In the revised manuscript, we have added a dedicated appendix section in Appendix A.1 to clarify the experimental setup and numerical procedures.

We now specify the prior distribution, the linear Gaussian measurement model, and the corresponding ground-truth posterior used as the reference. Specifically, the experiment studies the evolution of probability densities over a two-dimensional state space. The reported results are obtained by directly evolving and evaluating these densities on a discretized grid, rather than by drawing finite samples and estimating the densities. We also provide how DPS, DAPS, and A2PS are implemented in the density-level simulation, together with the grid configuration and relevant numerical settings. These additions clarify how the densities are propagated and evaluated, improving the reproducibility of the experiment.


Reference

[1] Song, Jiaming, Arash Vahdat, Morteza Mardani, and Jan Kautz. "Pseudoinverse-guided diffusion models for inverse problems." In International conference on learning representations. 2023.

[2] Boys, Benjamin, et al. "Tweedie moment projected diffusions for inverse problems." arXiv preprint arXiv:2310.06721 (2023).

[3] Hyungjin Chung, Jeongsol Kim, Michael Thompson Mccann, Marc Louis Klasky, & Jong Chul Ye (2023). Diffusion Posterior Sampling for General Noisy Inverse Problems. In The Eleventh International Conference on Learning Representations .

Dear authors: Thank you for providing a detailed response to my comments. I can now see a how this paper distinguishes itself from prior works such as PiGDM (i.e., W1 which was my main concern). Thank you also for providing additional experimental results. I will revise my score to reflect these changes.

Dear reviewer Gc11: Thank you for your thoughtful consideration and for taking the time to review our response. We sincerely appreciate your positive reassessment.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Summary of Reviewer Feedback and Author Responses

Dear Area Chair and Reviewers,

Thank you for your constructive feedback and detailed discussions. Throughout the rebuttal, we have clarified the theoretical contributions of A2PS, conducted additional experiments to validate its error-correction mechanism and posterior sampling quality, and expanded our analysis of baseline comparisons and experimental fairness. Below, we summarize the main strengths recognized in the reviews, the concerns raised, and the corresponding revisions and additional evidence. Links to the relevant point-by-point discussions are provided for further details.

1. Positive Aspects Highlighted in the Reviews

  • Importance of the problem: Reviewers recognized the importance of understanding error accumulation in approximate posterior guidance and its practical implications for inverse problems (Gc11, Jrdu).
  • Theoretical motivation: The asymptotic characterization of the endpoint conditional was recognized as mathematically interesting and providing principled motivation for posterior estimation (Gc11).
  • Methodological design: The proposed Langevin re-equilibration mechanism was considered conceptually well-motivated and potentially effective in addressing accumulated trajectory errors (Jrdu).
  • Intuitive experiments: The 2D toy experiments were recognized as informative and effective in illustrating the error-correction behavior of A2PS (kSTi).
  • Practical potential: The training-free nature of A2PS and its applicability to large generative backbones were also noted (kSTi).

2. Main Concerns in the Initial Reviews

  • Novelty and theoretical positioning: The relationship between A2PS and prior Gaussian conditional approximations, including Pi-GDM[1] and TMPD[2], required clarification. The distinction from optimization-based methods such as DiffPIR[3], NDTM[4], and RB-Modulation[5] also needed further explanation. (Gc11, kSTi)
  • Posterior sampling quality: LPIPS and FID alone were considered insufficient to establish posterior quality, motivating additional evaluations of posterior distributions and repeated samples. (Jrdu)
  • Experimental comparisons and fairness: Reviewers requested broader baseline comparisons, clearer hyperparameter configurations, additional reconstruction metrics, and more detailed computational cost evaluations. (Jrdu, kSTi)

3. Revisions and Additional Evidence

3.1. Theoretical contributions and relation to prior work.

We clarified that the Gaussian clean-conditional approximation used in Pi-GDM and TMPD is not itself the novelty of A2PS. Our contribution is to establish an asymptotic accuracy chain from the conditional approximation to posterior estimation and posterior-score estimation, with error orders of O(ρt3)O(\rho_t^3), O(ρt4)O(\rho_t^4), and O(ρt2)O(\rho_t^2), respectively. This analysis provides the theoretical foundation for using the increasingly accurate posterior score in Langevin re-equilibration to correct errors accumulated along the sampling trajectory.

We also clarified the distinction from DiffPIR, NDTM, and RB-Modulation. Although A2PS exhibits a guidance-balancing form, this formulation arises naturally from the surrogate posterior and surrogate likelihood constructed through our asymptotic analysis. Its derivation is motivated by asymptotically accurate posterior estimation, rather than by designing a mechanism to control the relative influence of measurement guidance and the generative prior. We expanded Sec. 3.2 and Sec. 4 to clarify these distinctions and discuss the relevant prior works.

(See detailed responses to Gc11 (novelty), and further response to kSTi regarding theoretical contributions and related work.)

3.2. 2D experiments: direct validation of the error-correction mechanism.

We expanded the 2D experiments beyond the original toy setting to cover nine configurations involving different prior distributions and measurement operators. These settings allow us to compute oracle posterior distributions and directly evaluate distributional discrepancies using TV and forward KL divergences.

The experiments distinguish the contribution of posterior-score estimation from that of Langevin re-equilibration: using the same score estimator without re-equilibration leads to accumulated distributional errors, whereas A2PS progressively reduces these errors toward the numerical integration floor. We further compare against DAPS [6], including a setting with exact full covariance, and conduct ablations on isotropic covariance and finite Langevin steps to examine the robustness of the mechanism under practical approximations.

These results provide direct evidence for the error-correction mechanism. They also extend the initial toy experiments that Reviewer kSTi recognized as intuitive and informative, and address Reviewer Jrdu's request for further validation of the proposed mechanism.

(See detailed responses to Gc11(Response to W3) and Jrdu on the additional 2D experiments and error-correction mechanism (Response to W2 and Q4).)

3.3. Image-level posterior sampling quality.

To evaluate posterior quality beyond conventional perceptual metrics, we introduced Energy Score [7] and pixel-wise marginal coverage on super-resolution, nonlinear deblurring, and checkerboard inpainting. The evaluation uses 50 images, two independent noisy observations per image, and eight posterior samples per observation, resulting in 800 generated samples per method per task.

A2PS achieves the lowest Energy Score among the compared methods across all three tasks. Its pixel-wise coverage values (0.770–0.802)(0.770–0.802) are also close to the theoretical reference of 7/9≈0.7787/9 \approx 0.778, while DPS [8] and DAPS exhibit substantially lower coverage in these evaluations. We additionally provide repeated-sample visualizations under fixed observations and different random seeds to illustrate the variability of generated posterior samples.

These metrics provide complementary evidence regarding image-level posterior sampling quality, beyond LPIPS and FID.

(See the detailed response to Jrdu on posterior quality, Energy Score, and coverage (3rd section).)

3.4. Broader baseline comparisons, reconstruction fidelity, and efficiency.

We added comparisons with Pi-GDM, DiffPIR, NDTM, and RB-Modulation, as well as fast-step comparisons with DDRM[9] and DDNM+[10]. The evaluation covers up to six ImageNet-256 inverse-problem tasks, using the recommended configurations from the respective papers or official implementations where applicable. The results show that A2PS with five LD steps achieves the best LPIPS on four of six tasks and the best FID on three of six tasks among the reported comparisons. Even with one LD step, A2PS achieves competitive perceptual quality at substantially lower runtime than several iterative baselines.

We also report PSNR and SSIM to provide complementary evidence on reconstruction fidelity and discuss the perception–distortion trade-off. Computational analysis includes NFE, per-image runtime, task-dependent LD costs, and one-time offline backbone profiling. The additional baseline results and settings are provided in Appendix H, while further qualitative comparisons are included in Fig. 6 and Appendix G.

(See detailed responses to kSTi on additional baselines, and to kSTi on runtime and reconstruction metrics.)

4. Additional Discussions with Reviewers

Reviewer Gc11. Following our clarification of the distinction between A2PS and prior work, particularly Pi-GDM, the reviewer explicitly acknowledged that their primary novelty concern had been addressed and indicated that they would revise their score. The score subsequently increased from 1 to 3.

Reviewer Jrdu. The reviewer acknowledged that the additional 2D experiments provided convincing evidence for the soundness of the proposed mechanism, while maintaining concerns about image-level comparisons and baseline tuning. We subsequently conducted additional DPS hyperparameter sensitivity experiments, clarified the metric used for hyperparameter selection, and added measurement-space error, PSNR, and SSIM evaluations for box inpainting. We also clarified that qualitative examples were manually selected for illustration and were not intended to represent a statistically representative subset of the test set.

(See Jrdu's follow-up discussions on qualitative comparisons and hyperparameter selection, and the subsequent response on measurement consistency.)

Reviewer kSTi. The reviewer requested further clarification of the theoretical contributions, the distinction from optimization-based methods, and the hyperparameter configurations of additional baselines. We expanded Sec. 3.2 and Sec. 4, clarified the motivation and derivation of the surrogate posterior formulation, and added the corresponding experimental settings and results in Appendix H. We also clarified that the additional baselines followed their published or official configurations without further task-specific hyperparameter searches.

(See kSTi's follow-up response and our subsequent clarification (starts from here).)

5. Overall

The review process has helped us sharpen the theoretical positioning of A2PS and substantially strengthen its empirical evaluation. In particular, the asymptotic analysis establishes the foundation for posterior re-equilibration, the 2D experiments directly validate its error-correction behavior against oracle posterior distributions, and the additional image-level evaluations and baseline comparisons provide broader evidence of its practical performance. We thank the reviewers for their constructive feedback, which has helped improve the theoretical clarity, experimental depth, and presentation of the work.

Reference

[1] Song et al., Pseudoinverse-Guided Diffusion Models for Inverse Problems. ICLR, 2023.

[2] Boys et al., Tweedie Moment Projected Diffusions for Inverse Problems. Transactions on Machine Learning Research, 2024

[3] Zhu et al., Denoising Diffusion Models for Plug-and-Play Image Restoration. CVPR Workshops, 2023.

[4] Pandey et al., Variational Control for Guidance in Diffusion Models. ICML, 2025.

[5] Rout et al., RB-Modulation: Training-Free Stylization using Reference-Based Modulation. ICLR, 2025.

[6] Zhang et al., Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing. CVPR, 2025.

[7] Gneiting et al., Strictly proper scoring rules, prediction, and estimation. Journal of the American statistical Association 102.477 (2007): 359-378.

[8] Chung et al., Diffusion Posterior Sampling for General Noisy Inverse Problems. ICLR, 2023.

[9] Kawar et al., Denoising Diffusion Restoration Models. NeurIPS, 2022.

[10] Wang et al., Zero-Shot Image Restoration Using Denoising Diffusion Null-Space Model. ICLR, 2023.

AuthorsOpenReview ↗

Correction to our previous response

We noticed two typographical errors in the experimental descriptions in our previous response.

  • Pi-GDM comparison (response to reviewer Gc11): We incorrectly stated that the comparison covered six tasks. The correct number is four tasks, as already reflected in the corresponding table.

  • Posterior sampling evaluation (response to reviewer Jrdu): We incorrectly stated the total number of generated samples as (800×2800 \times 2). The correct count is (800×3800 \times 3), corresponding to the three evaluated tasks. The reported table already contains the results from all three tasks.

These corrections concern the textual descriptions only; the corresponding experimental results and tables were already based on the correct settings. We apologize for these oversights.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.