acceptodds
Under review as a conference paper at ICLR 2027

Convergent Plug-and-Play Image Restoration with Annealed Noise Levels

Abstract

Plug-and-Play (PnP) methods solve imaging inverse problems by incorporating deep denoisers into iterative optimization algorithms. Although practical implementations often decrease the denoiser noise level along iterations, most existing convergence analyses assume a fixed denoiser. In this work, we establish convergence guarantees for a broad family of Plug-and-Play algorithms with annealed noise level, spanning deterministic methods (RED–GD and PnP–PGD) and stochastic methods (SNORE, equivariant RED, and a variant of PnP–Flow). For each method, we identify an explicit, nonconvex objective associated with the terminal denoising level and prove asymptotic stationarity of the iterates with respect to this objective. Our analysis does not prescribe any decay rate for the noise schedule, and our assumptions cover both learned gradient-step denoisers and exact MMSE denoisers. Overall, our theoretical results bridge the gap between existing PnP convergence theory and the decreasing-denoising practices used by state-of-the-art image restoration methods. We empirically demonstrate the benefits of such schedules and illustrate the predicted convergence behavior on several imaging inverse problems, including inpainting, super-resolution, demosaicing and tomography.

Primary area:
optimization
Keywords:
Plug-and-Play, Inverse Problems

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_tz8xOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the convergence of plug-and-play (PnP) image restoration algorithms under annealed denoising noise levels. The authors establish asymptotic stationarity guarantees for several deterministic and stochastic PnP variants, including RED-GD, PnP-PGD, SNORE, and ERED, under appropriate regularity and structural assumptions. Stronger convergence results are derived under additional conditions on the annealing schedule. The theoretical analysis is complemented by experiments on several image inverse problems, including inpainting, demosaicing, super-resolution, and tomography, demonstrating the potential benefits of noise annealing.

Critical strengths
  1. Understanding the convergence behavior of PnP algorithms under varying denoising noise levels is an important question, particularly because annealing is frequently used in practice without corresponding convergence guarantees. The paper provides a relatively systematic analysis across multiple deterministic and stochastic formulations, extending the theoretical understanding of these methods.

  2. The authors carefully distinguish asymptotic stationarity from stronger convergence guarantees and address nontrivial issues involving boundedness, stochastic updates, and the regularity of the underlying denoising operators. These technical considerations strengthen the theoretical contribution and go beyond a straightforward extension of fixed-noise convergence results.

Critical weaknesses
  1. The practical relevance of the theoretical guarantees remains insufficiently established due to restrictive and unverified assumptions.

The convergence analysis relies on strong global regularity and structural assumptions on the denoising operators and their associated objectives. However, the experiments employ learned denoisers whose compliance with these assumptions is not rigorously established. The empirical spectral regularization and finite-sample diagnostics used in the paper do not constitute guarantees that the required properties hold globally.

Although the authors discuss theoretical modifications intended to enforce certain regularity conditions, it remains unclear whether these modifications preserve all the structural properties required by the different convergence results or whether the modified operators correspond to those actually used in the experiments. This distinction is particularly important for the stochastic formulations, where the theoretical objective may depend on the denoiser's behavior beyond the region explored by finite experimental trajectories.

Consequently, the paper establishes convergence for a mathematically restricted class of operators, but does not convincingly demonstrate that its guarantees apply to the practical learned-denoiser algorithms used to motivate and validate the work. This substantially limits the practical implications of the theoretical results.

  1. The experimental evaluation does not directly validate the paper's central convergence claims or adequately isolate the effect of annealing.

The empirical evaluation primarily focuses on reconstruction quality, objective values, and differences between consecutive iterates. While these measurements are useful for assessing numerical behavior, they do not directly establish convergence to stationary points of the limiting objective, which is a central theoretical claim of the paper. In particular, diminishing update magnitudes can arise from decreasing step sizes without necessarily demonstrating the claimed stationarity.

Furthermore, the constant-noise and annealed configurations do not always optimize the same limiting objective, making it difficult to determine whether the observed improvements result from the annealing trajectory itself or from differences in the effective optimization objectives and hyperparameter configurations.

The experiments also predominantly employ rapidly decaying annealing schedules, leaving the broader theoretical conclusions concerning less restrictive annealing conditions largely unexplored. As a result, the empirical evidence demonstrates that annealing can improve reconstruction performance in certain settings, but provides limited direct validation of the distinctive convergence guarantees developed in the paper.

Questions for authors
  1. To what extent do the theoretical convergence guarantees actually apply to the learned denoisers used in the experiments? The required global regularity and structural assumptions are not established by finite-sample diagnostics or empirical spectral regularization. Could the authors clarify which theoretical assumptions are rigorously satisfied by the implemented algorithms, and whether the proposed regularity-enforcing modifications preserve the additional structural properties required by the convergence analysis?

  2. Can the authors provide direct empirical evidence supporting convergence to stationary points of the limiting objective? The current experiments mainly report reconstruction quality, objective values, and iterate differences, which do not directly verify the central stationarity guarantees. Could the authors evaluate appropriate stationarity residuals and demonstrate that the observed convergence behavior is consistent with the theoretical predictions, particularly for the stochastic variants?

  3. Can the authors disentangle the benefits of annealing from changes in the limiting objective and optimization configuration? Since the constant-noise and annealed settings do not always share the same final noise level and effective objective, the reported performance improvements cannot be attributed solely to the annealing trajectory. Could the authors provide comparisons under matched limiting objectives and clarify whether the corresponding hyperparameters were independently optimized?

5 comments

We thank the reviewer for recognizing the relevance of our analysis. We address each comment and question below. We have also uploaded a revised version of the manuscript, with changes in blue.

W1/Q1

The convergence analysis relies on strong global regularity and structural assumptions on the denoising operators and their associated objectives. However, the experiments employ learned denoisers whose compliance with these assumptions is not rigorously established. [...] The empirical spectral regularization and finite-sample diagnostics used in the paper do not constitute guarantees that the required properties hold globally.

First, we emphasize that our assumptions are sufficiently general to cover exact MMSE denoisers, unlike some existing convergence analyses of annealed PnP that rely on more restrictive assumptions, such as the bounded-denoiser condition of Chan et al. (2017) or the log-concavity assumption of Pesme et al. (2025). This is particularly important since trained denoisers are typically designed to approximate MMSE denoisers. We believe that establishing convergence guarantees in this setting is an important theoretical contribution. Beyond this, we explicitly design our trained denoisers to satisfy the structural assumptions of our convergence analysis, rather than relying solely on their approximation of MMSE denoisers.

More precisely:

For RED-GD and SNORE, the learned gradient-step denoiser Dσ=Id−∇gσD_\sigma=\mathrm{Id}-\nabla g_\sigma satisfies the assumptions required for convergence: Appendix A.1 establishes the required uniform Lipschitz continuity of ∇gσ\nabla g_\sigma and coercivity of gσg_\sigma, after regularization outside an arbitrarily large box B\mathcal B.

For PnP-PGD and SNOPnP, the denoiser must be a proximal operator. Our construction builds on Hurault et al. (2022). A sufficient condition for the gradient-step denoiser to be a proximal operator is Lg<1L_g<1, where LgL_g uniformly bounds the Lipschitz constants of ∇gσ\nabla g_\sigma. We use the following relaxation:

=\mathrm{Id}-\rho\nabla g_\sigma.$$ for which the sufficient condition becomes $\rho L_g<1$. Since $L_g$ is provably finite, **there always exists a sufficiently small $\rho>0$ satisfying this condition**. The soft spectral penalty complements this approach by encouraging smaller $L_g$, thereby reducing the need for strong relaxation. In practice, we estimate $L_g$ from Hessian spectral norms on validation images and select $\rho$ accordingly. These estimates do not certify a global bound on $L_g$, and hence do not rigorously guarantee proximality for the experimental choice of $\rho$. We explicitly acknowledge this limitation in conclusion and in Appendix C.1. > It remains unclear whether these modifications preserve all the structural properties required [...] or whether the modified operators correspond to those actually used in the experiments. We understand this concern as now referring to the regularization of the learned potential proposed in Appendix A.1, equation (17). This construction modifies the learned potential outside a prescribed box $\mathcal B$ to ensure global Lipschitz continuity of its gradient and coercivity. This modification preserves the gradient-step structure of the denoiser, and proximality can still be guaranteed through sufficiently strong relaxation, as explained above. Moreover, **the box $\mathcal B$ can be chosen arbitrarily large**. The regularization leaves the potential unchanged inside $\mathcal B$, and therefore does not affect denoiser evaluations occurring within this region. In practice, we observe that all deterministic and stochastic iterates of our experiments remain within $\mathcal B=[-10,10]^d$, so this choice of $\mathcal B$ do not change the iterates. In general, for Gaussian-perturbed inputs, the probability of leaving $\mathcal B$ during a finite run can also be made arbitrarily small by enlarging $\mathcal B$. > Could the authors clarify which theoretical assumptions are rigorously satisfied by the implemented algorithms, and whether the proposed regularity-enforcing modifications preserve the [...] properties required by the convergence analysis? To summarize, the Assumptions 3.3 for RED-GD and SNORE convergence is satisfied. For PnP algorithms, proximality of the denoiser can be guaranteed through a sufficiently small relaxation parameter. In practice, the relaxation parameter is selected from empirical spectral estimates.

W2/Q2

Can the authors provide direct empirical evidence supporting convergence to stationary points of the limiting objective?

We thank the reviewer for this suggestion. All the Figures now also report, for each algorithm, the normalized terminal stationarity residual

{\|\nabla F_\varepsilon(x_0)\|}$$ During the initial phase, when $\sigma_k\gg\varepsilon$, this gradient norm remains high, since the algorithm follows a different objective $F_{\sigma_k}$. Once $\sigma_k$ approaches $\varepsilon$, it decreases toward zero. This terminal gradient norms decrease in line with the theoretical best-iterate rates: $O(1/\sqrt{k})$ for the deterministic algorithms and $O\left(1/\sqrt{\sum_{j=0}^{k}\gamma_j}\right)$ for SNORE and ERED. This supports our stationarity results even when the stepsizes decrease: the gradient norms decrease too, so the small updates are not just a consequence of $\gamma_k\to0$. **Q3** > Can the authors disentangle the benefits of annealing from changes in the limiting objective and optimization configuration? [...]Could the authors provide comparisons under matched limiting objectives and clarify whether the corresponding hyperparameters were independently optimized? We thank the reviewer for this suggestion. In the original comparisons, the constant-noise and annealed configurations were independently tuned on the validation set to maximize PSNR. In the revised manuscript, the different Figures additionally compare annealing against constant $\sigma=\varepsilon$ (visual results and convergence curves). Both configurations share the same limiting objective but the constant-$\varepsilon$ configuration always produces worse reconstructions. For example on Figure 2, the large inpainting hole remains unfilled, and on Figure 3 substantial artifacts persist in demosaicing. These observations confirm the observation from introduction: constant small denoising level is insufficient to correct large scale degradations. Starting at a larger noise level with annealing enables to handle large structural degradations before reducing the denoising strength to recover finer details. > The experiments also predominantly employ rapidly decaying annealing schedules, leaving the broader theoretical conclusions concerning less restrictive annealing conditions largely unexplored. [...] We added in Appendix C.4 comparisons, on inpainting and demosaicing experiments, between the geometric schedule used in the paper and "polynomial schedules" $$\sigma_k=\epsilon+(\sigma_0-\epsilon)(1+k/k_0)^{-\beta}$$ These schedules were advocated for diffusion sampling by Karras et al (2022). The geometric schedule is recovered in the limit $\beta\to\infty$. For all these schedules, the terminal objective $F_\epsilon(x_k)$ levels off and $|\nabla F_\epsilon(x_k)|$ decreases, as predicted by our convergence results. Near the floor, the gradient norm decreases more slowly for smaller $\beta$. This is consistent with our convergence theory which proves $|\nabla F_{\sigma_k}(x_k)-\nabla F_\varepsilon(x_k)|=O(\sigma_k-\varepsilon)$. Concerning reconstruction quality, we observe that inverse problems with large missing regions, such as large-hole inpainting, need enough iterations at large $\sigma$: small $\beta$ fails to fill the hole, while larger $\beta$ matches the geometric schedule. For demosaicing, which does not have such a lare missing region, all schedules perform similarly. We thank the reviewer for their constructive feedback. We believe that the resulting revisions have significantly strengthened our paper and hope that our responses and additional results have addressed the reviewer's concerns. **References** Stanley H. Chan, Xiran Wang, and Omar A. Elgendy. Plug-and-play ADMM for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging. 2017. Samuel Hurault, Arthur Leclaire, and Nicolas Papadakis. Proximal denoiser for convergent plug-and-play optimization with nonconvex regularization. ICML 2022. Karras, Tero, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS 2022.

Thank you for the detailed responses and additional experiments. However, I believe several fundamental concerns remain unresolved.

  1. The gap between theoretical guarantees and practical PnP implementations remains substantial. The authors acknowledge that proximality in the experiments is assessed only through empirical Hessian spectral norms on validation data, which cannot establish the required global condition. The theoretical existence of a sufficiently small relaxation parameter does not imply that the experimentally chosen value (ρ=0.7\rho=0.7) satisfies this condition. Consequently, whether the PnP convergence guarantees actually apply to the implemented algorithms remains unestablished.

  2. The additional stationarity experiments do not fully justify the claimed theoretical convergence rates. The theoretical best-iterate rates generally concern the moving objective FσkF_{\sigma_k}, whereas the reported curves show the gradient norm of the terminal objective FϵF_\epsilon. These quantities are not interchangeable without additional error bounds. More importantly, for the stochastic variants, the theoretical rate bounds concern expected best-iterate gradients of the moving objective, while terminal stationarity is established asymptotically almost surely. Thus, the observed finite-iteration decay does not, by itself, establish the claimed rate behavior. The distinction is particularly important when interpreting decreasing update magnitudes in stochastic algorithms.

  3. The additional matched-objective experiments do not fully disentangle the benefits of annealing. Comparing annealing with constant σ=ϵ\sigma=\epsilon only demonstrates that starting directly with a small denoising level is ineffective. It does not establish that annealing outperforms the best fixed-noise configuration. Conversely, the independently tuned constant-σc\sigma_c baseline optimizes a different objective. Therefore, the experiments still do not cleanly isolate the contribution of the annealing trajectory from the effects of objective selection and hyperparameter tuning.

Overall, the revisions improve the experimental presentation, but they do not resolve the fundamental gap between the theoretical assumptions and the practical algorithms, nor do they fully establish the empirical claims made in support of the theoretical contributions. In particular, the global proximality required for the PnP convergence analysis remains unverified in the actual experiments. I therefore maintain my original rating of 2 (Weak Reject).

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

This paper provides convergence analysis for prior annealed Plug-and-Play (PnP) methods. Different variants of PnP algorithms such as PnP-GD, PnP-PGD, SNORE, and PnP-Flow are investigated. The authors provide convergence of objective function to stationary points using structured (proximal (MMSE) and gradient step) denoisers. They provide empirical evidence showing that generally using annealing noise level construct better solutions.

Critical strengths
  • This paper is a solid work that provides convergence result for annealed prior in five PnP framework, going beyond what exists in the literature on PnP annealing (Chan2017) and providing stationarity.

  • The authors provided empirical justification for some of the assumptions as well.

Critical weaknesses
  • Novelty is somehow narrow. From theoretical aspect, we have
    1. Annealing is a well stablished technique in various algorithms (Blake & Zisserman, 1987; Song et al., 2021; Dhariwal & Nichol, 2021) and it is already been put forward in PnP method as well (Zhang et al., 2017; Wei et al., 2020; Zhang et al., 2022; Zhu et al., 2023; Martin et al., 2025) and PnP convergence as well (Chan et al., 2017; Hou et al., 2022; Shrestha & Fu, 2026).

    2. Structured denoisers such as gradient step and proximal denoisers are well-investigated in the literature (Gribonval, 2011; Gribonval & Nikolova, 2020; Hurault et al., 2022a;b; Fang et al., 2024) and multiple existing works have looked into their convergence analysis (Ryu et al., 2019; Xu et al., 2020; Cohen et al., 2021; Hurault et al., 2022a;b; 2023; Renaud et al., 2024).

    3. The guarantee provided cannot separate the annealing from fixed noise. When we look at the Theorem 3.4, we see that the for objective function in annealed priors ∇Fε(xk)→0∇F_ε(x_k) \to0 and that every cluster point is a critical point of FεF_ε. We have the same result for RED–GD at a constant σ=εσ = ε (Hurault et al. (2022a)). While this makes sense since the total noise change is finite and can be construed that annealing is specific type of initialization, however, the algorithm converges from any start point. This point out that the purpose is showing that annealing does not hurt, though it might not help either. 

The main point of novelty is the specific technique used in theory where the monotically decreasing noise schedule has finite total variation which leads to look at the changes in regularization as summable perturbations.

  • The contribution is limited: Annealing has been shown to provide empirical gain before, and the same phenomenon is visible from the experimental results as well. The paper would been had a significant contribution if the theoretical results shed light on why the annealing provides limit point that are of better quality, or for example, extensive evidence that annealing provides a lower objective function.
  • Experimental result provided are somehow limited. Images only 128 x 128. No comparison with actual PnP-Flow + some diffusion based such as DiffPIR, or for example Pesme2025.
Questions for authors

Could the authors provide comparison of F for the cases of annealing vs. constant noise? ( Annealing is shown, but good to have constant as well) Could the authors also provide the objective F number for images shown? (Anneal final number is shown for example for Fig2, but no comparison with for example constant or even different constant sigmasigma)

Decoupling noise level from noise amplitude in the case of specifically PnP-flow creates mismatch for the flow trained at the certain noise level. Could the authors comment on the implication of this choice?

6 comments

We thank the reviewer for their constructive feedback and for acknowledging the soundness of our work. We address each comment and question below. We have also uploaded a revised version of the manuscript, with changes in blue.

Novelty is somehow narrow.

We believe our contribution addresses an important theoretical gap in the PnP literature. While decreasing noise schedules are widely used in state-of-the-art image restoration algorithms, most existing convergence analyses assume a fixed denoiser.

Annealing is a well stablished technique in various algorithms [...] and it is already been put forward in PnP method as well [...] and PnP convergence as well (Chan et al., 2017; Hou et al., 2022; Shrestha & Fu, 2026).

We do not claim to introduce annealing. Our contribution is to establish convergence guarantees for annealed PnP/RED algorithms under general decreasing noise schedules.

The cited convergence results address a different and narrower setting: fixed point convergence of PnP–ADMM, with very specific denoiser assumptions or algorithmic modifications. In particular:

  • (Chan et al., 2017) and (Hou et al., 2022) establish fixed-point convergence under the "bounded-denoiser" assumption. This assumption is not satisfied by the ideal MMSE denoiser, and thus it is unlikely to be verified by learned denoisers, which are built to approximate the MMSE.
  • (Shrestha and Fu, 2026) uses a special denoising procedure based on conditional Langevin dynamics. This is a specific and cumbersome change of the algorithms compared to general PnP methods.

In contrast, our result encompasses both MMSE denoisers and (gradient-step / proximal) learned denoisers. Our results additionally covers a wide variety of deterministic and stochastic PnP algorithms. Moreover, our convergence results are more informative as we show stationarity for an explicit objective instead of fixed-point convergence.

Structured denoisers such as gradient step and proximal denoisers are well-investigated in the literature [...] and multiple existing works have looked into their convergence analysis.

We agree that these denoisers are well-established in the PnP literature. Our theory first provides convergence guarantees for MMSE denoisers, which we believe constitutes an important theoretical contribution in itself. Beyond this, we explicitly use trained gradient-step and proximal denoisers to ensure that the structural assumptions of our convergence analysis are satisfied, rather than relying only on the approximation of MMSE denoisers, as common in the PnP literature.

The paper would been had a significant contribution if the theoretical results shed light on why the annealing provides limit point that are of better quality, or for example, extensive evidence that annealing provides a lower objective function.

We agree that our convergence results do not formally establish that annealing selects better solutions. However, proving convergence to stationary points of an explicit objective is a first non-trivial step to ensure stability and tractability of the iterates. We indeed prove convergence to the same objective FεF_\varepsilon than the algorithm performed with constant level σ=ε\sigma=\varepsilon. Annealing acts as a continuation strategy that influences which stationary point of FεF_\varepsilon is reached, and our experiments demonstrate that this can substantially improve reconstruction quality.

The reviewer suggests investigating if annealing achieves a lower terminal objective value . Interestingly, our experiments show that better reconstruction quality does not necessarily correspond to a lower value of FεF_\varepsilon. In the revised manuscript, all our Figures now additionally display, as dotted lines, the objective Fε(xk)F_\varepsilon(x_k) along the iterates of the same algorithms run with a constant level σ=ε\sigma=\varepsilon.

In particular, for large-hole inpainting (Figure 2) and sparse-view tomography (Figure 5), the fixed-noise method can achieve a lower terminal objective value while producing a visibly worse reconstruction. In large-hole inpainting, for instance, it may leave the missing region unfilled despite attaining a lower objective value. One possible explanation is that the regularizer gεg_\varepsilon, associated with a small denoising level, primarily captures fine-scale image statistics and may insufficiently penalize missing large-scale structures. Starting with stronger denoising can help recover these structures before progressively refining the reconstruction at smaller noise levels.

Finally, we agree that a suitable initialization could, in principle, enable the fixed-noise method to reach the same stationary point. However, finding such an initialization is challenging in high-dimensional imaging problems, if not impossible without using annealing strategies. Annealing offers a practical mechanism for guiding the optimization trajectory toward favorable stationary points within a single run.

Experimental result provided are somehow limited. Images only 128 x 128. No comparison with actual PnP-Flow + some diffusion based such as DiffPIR, or for example Pesme2025.

In the revised version, we provide Figure 7 a comparison on 512×512512\times512 natural images for random-pixel inpainting, showing consistent improvements from annealing. Although trained on smaller images, our GS-DRUNet and Prox-DRUNet denoisers use a fully convolutional architecture and can therefore be evaluated at larger resolutions.

Table 2 now additionally reports PSNR comparisons with the original PnP-Flow from Martin et al. (2025) and the diffusion-based method DiffPIR (Zhu et al. (2023)), which achieve restoration performance comparable to our annealed methods. In this controlled comparison, all methods use the same denoiser and are tuned on the same validation set. We do not include a direct comparison with Pesme (2025), whose annealing strategy is limited to the denoising problem.

Could the authors provide comparison of F for the cases of annealing vs. constant noise?

As detailed above, we now plot for each experiment the value of FεF_\varepsilon for each algorithm for both constant and annealed noise.

Decoupling noise level from noise amplitude in the case of specifically PnP-flow creates mismatch for the flow trained at the certain noise level. Could the authors comment on the implication of this choice?

Decoupling indeed introduces a mismatch between the denoiser training noise level and the input image. To control this mismatch, we use

+\varepsilon\left(1+\frac{k}{k_0}\right)^{-q}, \qquad q>1/2,$$ So $\tau_k$ and $\sigma_k$ initially remain close, then $\tau_k$ vanishes while $\sigma_k\to\varepsilon$. In particular, we always have \$ 0\leq\sigma_k-\tau_k\leq\varepsilon, \$ so the mismatch remains always small. This evolution enables theoretical stationarity without degrading reconstruction quality. The experiments indeed show that the denoiser tolerates this controlled discrepancy. Appendix Figures 4 and 5 of the revised manuscript additionally show the coupled variant, $\tau_k=\sigma_k$. In both cases, this variant retains a nonzero stationarity residual over the plotted iterations, whereas the decoupled variant approaches stationarity. The decoupled variant achieves better or comparable reconstruction quality. We thank the reviewer for their constructive feedback. We believe that the resulting revisions have significantly strengthened our paper and hope that our responses and additional results have satisfactorily addressed the reviewer's concerns. **References** Stanley H. Chan, Xiran Wang, and Omar A. Elgendy. Plug-and-play ADMM for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging. 2017. Ruizhi Hou, Fang Li, and Guixu Zhang. Truncated residual based plug-and-play ADMM algorithm for MRI reconstruction. IEEE Transactions on Computational Imaging. 2022. Rajesh Shrestha and Xiao Fu. Taming score-based denoisers in ADMM: A convergent plug-and-play framework. ICLR 2026. Segolene Martin, Anne Gagneux, Paul Hagemann, and Gabriele Steidl. Pnp-flow: Plug-and-play image restoration with flow matching. ICLR 2025. Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising diffusion models for plug-and-play image restoration. CVPR 2023.

Significance and novelty remain the two main concerns for this reviewer.

The analysis of annealed SNORE and Theorem 2 of SGPnP relate critical points to annealing in the harder limit σ→0\sigma\to0, where the target is the MAP objective; which the paper does not discuss. It seems that flooring σ\sigma at ε\varepsilon is what allows the stronger statement about the iterates in this paper. In Eqs. (37)–(41), annealed RED–GD is written as gradient descent on the fixed objective FεF_\varepsilon with an error ∥bk∥\|b_k\|. Stationarity then follows from the standard inexact descent inequality when this error is square-summable, and point convergence when it is summable. What does the general result add for schedules used in practice, and does the same reduction apply to the stochastic algorithms?

Significance: The guarantee for annealing is the same as for a constant σ=ε\sigma=\varepsilon, and the revised experiments show that FεF_\varepsilon is not consistent reconstruction quality (Figure 5). The theory therefore does not explain the benefit of annealing that motivates the paper.

Reviewer_rokGOpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper establishes convergence guarantees for several RED and PnP algorithms with denoising levels decreasing to a positive floor. Under structured denoiser assumptions, it proves stationarity for explicit terminal objectives, with point convergence for deterministic methods under faster annealing. Experiments on faces and natural images across four inverse problems demonstrate PSNR gains over constant noise baselines.

Critical strengths
  • Useful extension of convergence theory. Establishing terminal stationarity under any monotone noise schedule approaching a positive floor addresses a practical limitation of fixed denoiser analyses. For SNORE, this goes beyond the earlier analysis of critical point sets under annealing to establish convergence properties of the changing-noise iterates.

  • Consistent empirical support for annealing. Tables 2 and 4 show improvements across five algorithms and multiple inverse problems. Some gains are substantial: RED–GD improves from 33.54 to 37.89 dB on face demosaicing. These comparisons substantiate the practical relevance of the theoretical question.

Critical weaknesses
  • The connection between the guarantees and experimental denoisers remains conditional. Assumptions 3.3 and 3.8 require uniform regularity and, for learned proximal denoisers, contraction bounds. The experiments use sampled Hessian estimates and a soft spectral penalty followed by relaxation (Appendix C.1). These provide evidence but do not establish the required bounds. Furthermore, the cutoff construction ensuring global regularity can change the Lipschitz constant governing admissible stepsizes. Consequently, the claim that the reported implementations satisfy the convergence conditions is insufficiently supported.

  • The convergence plots do not directly measure stationarity. Figures 2–6 show objective stabilization and shrinking iterate differences. For SNORE and ERED, shrinking stepsizes can themselves produce small iterate differences. Without terminal gradient norms or another stationarity residual, these plots provide incomplete validation of the main theoretical conclusion.

Questions for authors
  • What evidence supports the required contraction and Lipschitz bounds for the actual experimental denoisers and stepsizes, including any cutoff transition region?

  • Do terminal objective gradient norms decrease in the stochastic experiments? Can the authors distinguish stationarity from small updates caused by vanishing stepsizes?

4 comments

We thank the reviewer for recognizing our contributions and for the constructive questions. We address each comment and question below. We have also uploaded a revised version of the manuscript, with changes in blue.

Assumptions 3.3 and 3.8 require uniform regularity and, for learned proximal denoisers, contraction bounds. The experiments use sampled Hessian estimates and a soft spectral penalty followed by relaxation.

First, we emphasize that our main contribution is to establish convergence of annealed PnP algorithms to stationary points of explicit objectives. We believe that establishing such guarantees for MMSE denoisers constitutes an important theoretical contribution in itself. We go further by explicitly designing our trained denoisers to satisfy the structural assumptions required by our convergence analysis (beyond the fact that they approximate the MMSE) which is rarely made in the PnP literature.

For convergence of RED-GD and SNORE, the learned gradient-step denoiser Dσ=Id−∇gσD_\sigma=\mathrm{Id}-\nabla g_\sigma satisfies the assumptions required for convergence: Appendix A.1 establishes the required uniform Lipschitz continuity of ∇gσ\nabla g_\sigma and coercivity of gσg_\sigma, after adding a regularization term outside an arbitrarily large ball B\mathcal B.

For convergence of PnP-PGD and SNOPnP, the denoiser must be a proximal operator. Designing proximal neural networks is a separate research direction, which has recently been extensively studied (e.g. Hurault et al. (2022), Fang et al. (2024), Pesquet et al. (2020)). Our construction builds on Hurault et al. (2022). A sufficient condition for the gradient-step denoiser to be a proximal operator is Lg<1L_g<1, where LgL_g uniformly bounds the Lipschitz constants of ∇gσ\nabla g_\sigma. We use the following relaxation:

=\mathrm{Id}-\rho\nabla g_\sigma.$$ for which the sufficient condition becomes $\rho L_g<1$. Since $L_g$ is provably finite, **there always exists a sufficiently small $\rho>0$ satisfying this condition**. The soft spectral penalty complements this approach by encouraging smaller $L_g$, thereby reducing the need for strong relaxation. In practice, we indeed need to estimate $L_g$ from the Hessian spectral norm of $g_\sigma$ on validation images and select $\rho$ accordingly. We acknowledge that these empirical estimates do not provide a certified global upper bound on $L_g$ and thus cannot guarantee that the resulting denoiser is globally a proximal operator. We explicitly acknowledge this limitation in Appendix C.1 and in the conclusion of the revised manuscript. > The cutoff construction ensuring global regularity can change the Lipschitz constant governing admissible stepsizes. The reviewer refers to the additive regularization introduced in Appendix A.1, Eq. (17), which modifies the learned potential outside a prescribed box $\mathcal B$ to ensure global Lipschitz continuity of its gradient and coercivity. We agree that this construction may change the global Lipschitz constant of the potential. However, **the box $\mathcal B$ can be chosen arbitrarily large**. The regularization leaves the potential unchanged inside $\mathcal B$ and thus does not affect the algorithm as long as as long as all inputs passed to the denoiser remain inside this box. In practice, we observe that all deterministic and stochastic iterates remain within $\mathcal B=[-10,10]^d$. For Gaussian perturbations, the probability of leaving the box during a finite run can also be made arbitrarily small by enlarging $\mathcal B$. Since the stepsizes are chosen based on empirical estimates of the Lipschitz constant, the cutoff does not affect this empirical choice. Second, for deterministic algorithms, an alternative to avoid this regularization is to explicitly assume boundedness of the iterates, as commonly done in nonconvex optimization (Attouch et al. (2013), Hurault et al. (2021)). Joint $\mathcal C^2$ regularity of $g_\sigma$ then ensures uniform Lipschitzness of $\nabla g_\sigma$ on any compact convex set containing the iterates. The same convergence analysis then follows using these local bounds and the corresponding stepsize conditions. For RED–GD, this boundedness can also be enforced by adding an explicit projection onto a compact convex set, yielding a projected GD variant. Similar localization argument could also be adapted to stochastic algorithms (SNORE and SNOPnP) but these extensions are beyond the scope of the present response. We have clarified these limitations and alternatives in Appendices A.1 and C.1 of the revised manuscript.

The convergence plots do not directly measure stationarity. [...] Do terminal objective gradient norms decrease in the stochastic experiments? Can the authors distinguish stationarity from small updates caused by vanishing stepsizes?

We thank the reviewer for this suggestion. Figures 2–6 now report, for each algorithm, the normalized terminal stationarity residual

{\|\nabla F_\varepsilon(x_0)\|}$$ During the initial phase, when $\sigma_k\gg\varepsilon$, this gradient norm remains high, since the algorithm follows a different objective $F_{\sigma_k}$. Once $\sigma_k$ approaches $\varepsilon$, it decreases toward zero. This terminal gradient norms decrease in line with the theoretical best-iterate rates: $O(1/\sqrt{k})$ for the deterministic algorithms and $O \left(1/\sqrt{\sum_{j=0}^{k}\gamma_j}\right)$ for SNORE and ERED. This supports our theoretical results even when the stepsizes decrease: the gradient norms decrease too, so the small updates are not just a consequence of $\gamma_k\to0$. We thank the reviewer for their constructive feedback. We believe that the resulting revisions have significantly strengthened our paper and hope that our responses and additional results have addressed the reviewer's concerns. **Reference** Attouch, Hedy, Jérôme Bolte, and Benar Fux Svaiter. Convergence of descent methods for semi-algebraic and tame problems: proximal algorithms, forward–backward splitting, and regularized Gauss–Seidel methods. 2013. Fang, Zhenghan, Sam Buchanan, and Jeremias Sulam. What's in a prior? Learned proximal networks for inverse problems. ICLR 2024. Samuel Hurault, Arthur Leclaire, and Nicolas Papadakis. Gradient step denoiser for convergent plug-and-play. ICLR 2022. Samuel Hurault, Arthur Leclaire, and Nicolas Papadakis. Proximal denoiser for convergent plug-and-play optimization with nonconvex regularization. ICML 2022. Pesquet, Jean-Christophe, Audrey Repetti, Matthieu Terris, and Yves Wiaux. Learning maximally monotone operators for image recovery. SIAM Journal on Imaging Sciences. 2021.
Reviewer_rokGOpenReview ↗

Response to the rebuttel

Dear authors,

Thanks for the detailed response and for updating the paper.

The stationarity plots added address my concern about small updates caused by decreasing stepsizes. I also appreciate the clearer discussion of the assumptions and their limitations.

My remaining concern is that the required bounds still haven’t been verified for the denoisers and stepsizes used in the experiments. The cutoff and localization arguments help clarify the situation, but this gap remains. Therefore, I I regard this as a limitation of the theory’s applicability to the implementations and recognize the theoretical results useful under the stated assumptions. I’m keeping my current recommendation.

Thank you for the follow-up. We appreciate your confirmation that the new stationarity plots address the concern about decreasing step sizes.

Global proximality is required by the convergence guarantees for PnP–PGD and SNOPnP. The analyses of RED–GD, SNORE, and ERED use gradient-step denoisers. To evaluate both PnP methods with a globally certified proximal denoiser instead of an empirically verified one, we added a demosaicing experiment in Appendix C.7. We trained a denoiser parameterized as the gradient of a smooth input-convex neural network (ICNN), which makes it exactly proximal by construction Fang et al., ICLR 2024

On 100 FFHQ test images, annealing still improve PSNR over fixed noise level tested by 1.84 dB for PnP–PGD and 1.82 dB for SNOPnP. We also assessed convergence numerically for both annealed methods.

Comments on the paper

1 comment

We thank the reviewers for the constructive discussion. We summarize the main revisions below, including a new experiment addressing the shared concern about global proximality.

Exactly proximal denoisers (rokG, tz8x). Of the five methods studied, the convergence guarantees for PnP–PGD and SNOPnP require proximal denoisers. The empirical spectral estimates leave global proximality uncertified for the learned denoisers used in the original experiments. In Appendix C.7, we now evaluate both PnP methods on demosaicing with a denoiser parameterized as the gradient of a smooth input-convex neural network, which ensures exact proximality by construction. On demosaicing, annealing still improve mean PSNR by more than 1.8 dB for both methods over the best fixed noise level tested. Convergence is verified numerically in Figure 12.

Scope of the theory (pzQc). We clarify that our moving-objective analysis establishes asymptotic stationarity when the denoising level decreases at every iteration with no prescribed decay rate. We added in Remark 3.5 and Appendix B.1.2 discussions about the additional regularity needed for using a zero noise floor.

Stationarity and reconstruction (rokG, pzQc, tz8x). Figures 2–6 now report terminal-objective gradient norms. Figure 11 compares moving- and terminal-objective gradient norms. The revision also adds fixed-noise comparisons with the same terminal objective, a single-switch warm-start ablation, 512 × 512 image inpainting results, and comparisons with PnP-Flow and DiffPIR.

We believe these revisions address the weaknesses and questions raised in the reviews and substantially strengthen the manuscript. Thank you again for the feedback that guided this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.