acceptodds
Under review as a conference paper at ICLR 2027

MAP-convergent Denoising in Nonconvex PnP for Inverse Problems

Abstract

Plug-and-Play (PnP) has become an increasingly popular framework for inverse problems, enabling optimization formulations with learning-based priors in proximal splitting methods. Score-based generative models have recently emerged as strong learned priors, but it remains challenging to efficiently utilize them in splitting-based PnP frameworks. Specifically, PnP frameworks often use denoisers as implicit priors in algorithms motivated by Maximum a Posteriori (MAP) estimation, but the annealed nature of score-based denoisers makes the theoretical connection between MAP and PnP challenging. Another difficulty comes from the geometry of the optimization iterates, which can deviate significantly from the noisy data that the score-based model is trained on. We consider MAP estimation with score-based PnP through primal-dual splitting and show that under suitable conditions, the primal-dual iterates converge to the stationary set of the corresponding nonconvex and nonsmooth MAP objective function. Additionally, we quantify the geometry mismatch introduced by the splitting method, while also revealing a tradeoff between convergence and geometrical alignment of the iterates. Our primal-dual approach yields updates which admit efficient closed-form solutions for structured nonlinear problems, including phase retrieval and sparse-view transmission computed tomography. Experiments on linear and nonlinear inverse problems show competitive reconstruction quality with fewer score evaluations and shorter runtimes compared to the state-of-the-art.

Primary area:
optimization
Keywords:
Optimization, Diffusion Models, Score-based Generative Models, Plug-and-Play

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_6MB9OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3(+2)
Soundness
3(+1)
Read review
Paper summary

This work considers primal dual Plug-and-Play methods, which replace some regularization term in an operator splitting with a learned denoisers. The work is built from PDHG, the denoiser is assumed to be a Gaussian denoiser with variable noise, with the proposed algorithm injecting additional noise at each step. The authors then provide two theoretical results. The first shows convergence to critical points of some functional when the injected noise is decreasing, under standard assumptions in the literature, namely extending(?) previous analysis to the nonconvex case. The second then provides a recovery guarantee when the added noise is the same as the denoising strength. The proposed method is then compared with several diffusion based solvers.

Critical strengths
  1. Theory is reasonable and addresses different choices of injected noise. The main contribution Theorem 2 actually appears to be a recovery guarantee, in the specific case where the denoiser is given by a MMSE denoiser (with decreasing noise levels). This actually has interesting links to signal recovery [LT26]. Other similar works in the literature typically consider recovery under diffusion processes, rather than such iterative procedures.

[LT26] O. Leong and Y. Traonmilin. A Recovery Theory for Diffusion Priors: Deterministic Analysis of the Implicit Prior Algorithm. AISTATS'26

Critical weaknesses
  1. Clarity.
    • The introduction does not serve to position this work relative to other related work, and reads instead like a general introduction to PnP. There are also several missing references.
    • Section 2 is particularly difficult to read. Is there any particular reason why the proximal interpretation is introduced here? It is well known within the PnP literature that fixed points of many algorithms can be written as critical points of some functional, related to MAP estimates through the Bayesian interpretation. It would be much better to detail the differences of the proposed method (i.e. added noise at each step + decreasing denoiser variance) with more classical PnP methods using fixed denoisers. There should also be a review on diffusion-based PnP.
    • Please rearrange the text so that the writing flows naturally, rather than having abrupt bold text in front of a paragraph. E.g. "Manifold hypothesis" in Section 4.1,
  2. Experimental section. Are the baselines tuned? I could not find an explanation in the appendix. It is well known that PnP (and by extension, RED) methods are sensitive to hyperparameters and stopping time.
    • I would also be interested in Figure 2 but with the other compared methods as well, e.g. DiffPIR, which does not have to do potentially expensive dual iterations.
Questions for authors
  • Are the sks^k in Theorem 2 equal to limit points of the PD algorithm? Can anything else be said about them, other than they sit on the data manifold? I suspect that this may be difficult since the score explodes as the variance goes to 0. This is related to Appendix G, uniform bound on δk\delta_k in Prop 5 would not be feasible as variance goes to 0 in practice.
  • Please clarify usage of "MAP estimate" in the case where your objective is nonconvex.
  • Metrics increase but do not converge in the experimental section. Any demonstration of Theorem 1 convergence to fixed points?
  • Consider rewriting Section 3.2 to have more motivation. It currently reads rather weakly as "PDHG -> replace a denoiser like PnP -> add noise because the literature says so".
  • (Suggestion) Proposition 1 should probably highlight that even though the dual problem (8) is not equivalent to (2) if hh is nonconvex, if the primal iterations converge, then it converges to a critical point of the primal problem (2).
  • (Suggestion) Visually demonstrate the variety of images that can be reconstructed from a given corruption. I see that Appendix K has statistics.

Missing references for PnP-PDHG should be added, e.g. [MMHC17] and associated works by Ono [Ono17].

[MMHC17] Meinhardt, T., Moller, M., Hazirbas, C., & Cremers, D. (2017). Learning proximal operators: Using denoising networks for regularizing inverse imaging problems. In Proceedings of the IEEE international conference on computer vision (pp. 1781-1790). [Ono17] Ono, S. (2017). Primal-dual plug-and-play image restoration. IEEE Signal Processing Letters, 24(8), 1108-1112.

Reviewer_GZL1OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a primal-dual plug-and-play method using MMSE denoisers with decreasing denoising levels. Its main result establishes almost-sure convergence to the stationary set of the original MAP objective under a Lipschitz-gradient assumption on the prior and specific noise and stepsize schedules. Experiments cover several linear and nonlinear inverse problems, including phase retrieval and transmission CT.

Critical strengths
  1. In my view, the central contribution is the analysis of an annealed PnP iteration that approaches stationarity of the original MAP objective as the denoising level tends to zero. This deserves to be foregrounded more explicitly in the abstract and contribution list.

  2. Proving nonconvex convergence of PnP-PDHG without assuming a smooth data fidelity term is new and technically interesting. It makes it possible to use convergent Plug-and-Play for Poisson inverse problems with denoising priors. This should be emphasized in the introduction.

  3. The inclusion of linear restoration tasks, phase retrieval, and transmission CT gives a broad numerical picture. Reporting wall-clock performance alongside score-evaluation counts is also useful.

Critical weaknesses

1. MAP vs MMSE motivation

The statement that "replacing a proximal prior update with a Tweedie denoiser introduces a gap between MMSE denoising and MAP estimation" is too ambiguous. My understanding is that the intended issue is the following: the proximal map of the original negative log-prior is a MAP denoiser, whereas the trained network approximates the MMSE denoiser for that prior, and these two maps generally differ. This should be stated directly, including in the abstract.

An important missing qualification is that the exact Gaussian MMSE denoiser is itself a proximal operator of another, noise-dependent potential. This is established by Gribonval (2011), with related consequences for nonconvex PnP optimization developed by Hurault, Leclaire, and Papadakis (2022). Thus, the issue is not the absence of a variational interpretation, but its relation to the original prior and its evolution as the noise level changes. Writing the MMSE denoiser as a proximal map of could also provide a clearer proof of Proposition 2 and the annealing argument.

The discussion should be clear that there is a whole family of plug-and-play methods which replace the score by a denoising residual (sometimes called RED) that do target the right MAP problem.

2. Literature comparison

Table 1 should distinguish convergence at a fixed denoising level from convergence along an annealed trajectory. SNORE targets a relaxed MAP objective at fixed σ\sigma, but already establishes consistency with the original MAP problem as σ→0\sigma \to 0 and proposes an annealed algorithm. The present paper’s additional contribution is to prove convergence of a single trajectory with σk→0\sigma_k \to 0. Thus, the distinction is more precise than “relaxed MAP versus MAP”: it concerns the convergence guarantee when the denoising level changes during optimization.

The comparison with Sun et al. should also be more precise. In Sun et al. (2018), Lemma 3.4 already gives cluster-point stationarity for the exact proximal iteration without invoking KL; the stronger whole-sequence convergence theorem uses semi-algebraicity/KL. Consequently, avoiding KL while proving approach to a stationary set is not, by itself, the distinguishing advance.

3. Choice of primal-dual splitting. It remains unclear how much of the practical benefit comes from the choice of PDHG rather than from annealing, noise injection, or parameter tuning. A comparison with PnP-ADMM and/or PGD using the same pretrained denoiser and comparable tuning and evaluation budgets would be informative, particularly on the smooth linear tasks. Comparing fixed and decreasing denoising levels would further isolate the benefit of annealing.

4. Theorem 2 Theorem 2 shows a tension between denoiser-input alignment and MAP convergence: matched noise ρk=σk\rho_k=\sigma_k gives alignment as sigmak→0sigma_k\to0 but is incompatible with the sufficient convergence conditions of Theorem 1. However, it remains unclear how this observation guides the practical choice of ρk\rho_k. Appendix J.2 specifies a two-stage schedule, with matched noise initially and a different decay for the final 20 iterations, but the choice of decay and switching time is not justified. In particular, there is no empirical comparison with either ρk=σk\rho_k=\sigma_k throughout or ρk≡0\rho_k\equiv0 even though the latter already satisfies the theorem's noise condition.

5. The experiments do not illustrate convergence.

There are no numerical curves for a stationarity measure, primal/dual residuals, or iterate increments. Reconstruction metrics versus runtime or iteration budget do not establish the behavior predicted by Theorem 1. If the original MAP objective cannot be evaluated for the pretrained model, computable quantities such as ∥xk+1−xk∥\|x^{k+1}-x^k\| ∥Axk−zk+1∥\|Ax^k-z^{k+1}\| and ∥wk+1−wk∥\|w^{k+1}-w^k\| would be informative.

Appendix J.2 uses the theorem-compatible schedule only for the final 20 iterations. Such a short tail does not demonstrate that the reported reconstructions approach the stationary set. A longer run with a fixed initial stage would help assess this.

6. Learned denoiser Appendix G does not extend the full conclusion of Theorem 1: it guarantees only the existence of one approximately stationary accumulation point, rather than convergence to an approximately stationary set or convergence of objective values. Moreover, its uniform score-error bound is neither verified nor quantitatively estimated for the experimental models. The guarantee relevant to the experiments is therefore substantially weaker and depends on an assumption whose practical validity remains unclear.

Questions for authors

See above

Reviewer_kCLmOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a primal-dual plug-and-play (PnP) framework for solving inverse problems using annealed score-based denoisers. The authors establish convergence to the stationary set of the original MAP objective under assumptions allowing nonconvex and nonsmooth data-fidelity terms. They further analyze the mismatch between denoiser inputs and Gaussian-corrupted training data, identifying a tradeoff between noise-scale alignment and convergence. Experiments on several linear and nonlinear inverse problems demonstrate competitive reconstruction quality with fewer score evaluations and reduced computational costs.

Critical strengths

The paper is generally clearly written

Critical weaknesses

Convergence analysis of nonconvex PnP algorithms has already been extensively investigated. For example, Hurault et al. (ICML 2022) established convergence to stationary points of an explicit objective for PnP algorithms with nonconvex regularization, including settings with nonconvex and nonsmooth data-fidelity terms. Park et al. (CAMSAP 2023) also investigated the convergence of nonconvex PnP-ADMM using MMSE denoisers. While the present paper considers annealed score-based denoisers and convergence to the original MAP objective, rather than a denoiser-induced regularized objective, the underlying approach appears to be a relatively incremental extension of existing results. In particular, the approximation of the MMSE denoiser by a proximal operator in the small-noise regime, combined with diminishing stepsizes and standard convergence arguments, seems conceptually natural. I would appreciate a more precise explanation of which theoretical difficulties are fundamentally new and cannot be addressed using existing nonconvex PnP convergence frameworks. The related-work discussion should also more carefully distinguish the proposed contribution from prior results.

Theorem 1 relies on asymptotic conditions on the denoising noise level, stepsize, and injected-noise amplitude. However, the experiments use a two-stage schedule in which most iterations employ matched noise levels and a fixed primal stepsize, followed by only 20 iterations using the theoretically motivated schedule. While this does not contradict the asymptotic theorem, it remains unclear how much the theoretical convergence mechanism contributes to the observed reconstruction performance. An ablation comparing different noise-injection schedules, including matched noise throughout, no additional noise injection, and different lengths of the final convergence stage, would help establish the practical relevance of the theoretical analysis. It would also be useful to report a measure of MAP stationarity alongside reconstruction quality.

Questions for authors

What is the essential theoretical novelty of Theorem 1 compared with existing nonconvex PnP convergence results, particularly Hurault et al. (2022) and related MMSE-denoiser analyses?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.