- 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
-
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.
-
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
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 ↑
| Method | Box Inpainting | Checkerboard Inpainting | Random 70% Inpainting | Super Resolution | Gaussian Deblurring | Non-linear Deblurring |
|---|---|---|---|---|---|---|
| piGDM | 16.96 / .641 | 21.98 / .607 | 23.56 / .587 | 23.18 / .548 | *(1) | *(1) |
| DiffPIR | 19.31 / .668 | 23.82 / .693 | 25.51 / .693 | 24.30 / .627 | 22.30 / .421 | *(1) |
| NDTM | 15.47 / .724 | 17.81 / .637 | 26.19 / .773 | 21.99 / .467 | *(1) | 24.41 / .657 |
| RB-Modulation | 16.76 / .625 | 17.34 / .518 | 25.40 / .723 | 22.81 / .521 | *(1) | 22.85 / .579 |
| A2PS (Ours) | 19.82 / .722 | 23.40 / .701 | 25.11 / .700 | 22.65 / .509 | 23.43 / .549 | 24.28 / .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)
| Method | NFEs | Box Inpainting | Checkerboard Inpainting | Random 70% Inpainting | Super Resolution | Gaussian 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.
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) to estimate the ideal posterior score st∣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∣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→1: \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 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 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)=btσ02at2∂x∂μ1∣t(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):
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