acceptodds
Under review as a conference paper at ICLR 2027

Metropolis-SSD: Symmetric Stochastic Denoisers for Convergent Image Reconstruction

Abstract

Denoiser-driven methods have become a powerful approach to iterative image reconstruction. Deep denoisers can achieve excellent reconstruction quality, but convergence of the resulting iterations is difficult to guarantee. Such guarantees typically require global properties such as nonexpansiveness, which are difficult to enforce for nonlinear denoisers. Linear denoisers offer a more tractable alternative, as their spectral properties can be analyzed and controlled directly. Their limited expressive power, however, can restrict reconstruction quality. A recent approach, NECTR, bridges this gap by separating nonlinear weight prediction from linear pixel aggregation, using a pretrained network to predict the weights. For quadratic data fidelity, NECTR guarantees convergence by enforcing the resulting linear denoiser ensuring that the resulting linear denoiser is stochasticity can easily be enforced by normalizing the predicted weights, imposing symmetry simultaneously is more challenging. We introduce Metro-SSD, which uses a simple Metropolis-type normalization to enforce both symmetry and stochasticity. Metro-SSD admits an exact matrix free implementation requiring only two aggregation passes, making it practical for large scale problems. Under a mild condition on the imaging forward model, we establish convergence guarantees for Metro-SSD with PnP-HQS and PnP-DRS. Using an attention-based network to predict the aggregation weights, Metro-SSD outperforms or matches the performance of state-of-the-art baselines in deblurring, superresolution and MRI.

Primary area:
optimization
Keywords:
image reconstruction, regularization, inverse problems, denoising, iterative methods, convergence

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

The present paper proposes Metro-SSD, a trainable denoiser for plug-and-play image reconstruction. A network predicts aggregation weights from a reference image. These weights define a symmetric operator, and Metropolis normalization makes the operator stochastic without forming the full matrix. After warm-up, the reference and algorithm parameters are fixed. For quadratic data fidelity, the authors establish linear convergence under the stated assumptions for PnP-HQS and PnP-DRS. DRS uses an equal average of the denoiser and the identity operator. Experiments on deblurring, superresolution, and MRI show competitive reconstruction performance.

Critical strengths
  • The construction enforces symmetry and stochasticity by design. The convergence analysis establishes linear convergence under the stated assumptions once the reference and algorithm parameters are fixed. This provides an explicit convergence guarantee while allowing flexibility in the neural weight predictor.

  • The method gives competitive results across deblurring, superresolution, and MRI. Compared with NECTR, the average PSNR improves by 0.78 to 0.85 dB for deblurring and 0.58 dB for superresolution. These results support the practical value of the complete method.

Critical weaknesses
  • The claimed advantages over NECTR are not fully established. The paper suggests that NECTR requires positive semidefiniteness, but NECTR's convergence analysis already permits negative eigenvalues. After computing the degrees, NECTR can also accumulate the normalized operator output and its row sums in the same pass, using the same type of combined accumulation as Metro-SSD. The stated two-pass versus three-pass comparison therefore does not establish an inherent computational advantage.

  • The experiments do not isolate the contribution of Metropolis normalization. The implementation combines a DRUNet-based attention predictor, direct training of the normalized denoiser, and a modified warm-up procedure. The ablations examine warm-up choices, but do not compare the two normalizations with the same predictor, training, and reconstruction settings. Therefore, the gains over NECTR support the complete method, while it remains unclear how much of the improvement comes from the proposed normalization.

Questions for authors
  1. What is the precise difference between the admissible denoiser classes of Metro-SSD and NECTR?

  2. How do runtime and memory usage compare when both methods use the same fusion and caching strategy?

  3. Can the authors compare the two normalizations with the predictor, training, and warm-up settings held fixed to clarify the source of the reconstruction gains?

  4. For the channelwise implementation, how does the convergence analysis exclude nonzero images that are constant within each channel but produce zero measurements? Explain how the RGB and complex MRI representations satisfy the required condition.

3 comments

(C1) Ablation of normalization and training. Appendix E.3.1 compares three variants using the same DRUNet-based attention predictor architecture, alongside NECTR-Proj with the original predictor:

MethodNormalization in trainingNormalization in reconstruction
NECTR-Proj (original predictor)Row normalizationNECTR
NECTR-AttnRow normalizationNECTR
NECTR-Attn-TrNECTRNECTR
Metro-SSDMetropolisMetropolis

NECTR-Proj versus NECTR-Attn evaluates the predictor change while retaining the normalization protocol. NECTR-Attn versus NECTR-Attn-Tr compares post-hoc and direct training while retaining NECTR normalization at reconstruction. NECTR-Attn-Tr versus Metro-SSD compares the two normalizations when each is used during both training and reconstruction.

The mean PSNR values below use the same test sets and task-level reconstruction settings within each row: PnP-HQS on CBSD10 and PnP-DRS on the first ten knee MRI test images for each reported setting. Appendix E.2 studies warm-up separately.

ApplicationSettingNECTR-ProjNECTR-AttnNECTR-Attn-TrMetro-SSD
DeblurringKernel 1, ν=0.03\nu=0.0328.2328.4329.1929.23
DeblurringKernel 6, ν=0.05\nu=0.0527.4727.5228.3728.42
DeblurringKernel 8, ν=0.05\nu=0.0526.6226.7827.0527.05
Superresolutions=2s=2, kernel 1, ν=0.05\nu=0.0526.6127.2127.5327.66
Superresolutions=3s=3, kernel 3, ν=0.03\nu=0.0325.1125.3925.8525.80
MRI2-fold single-coil PD43.1044.0644.0044.05
MRI4-fold multi-coil PDFS41.0241.0241.2441.28

The attention predictor improves over NECTR-Proj in six settings and ties in the seventh. Direct NECTR training improves over post-hoc normalization by 0.27–0.85 dB across the five deblurring/SR settings and by 0.22 dB on multi-coil MRI, with a 0.06 dB decrease on single-coil MRI. Metro-SSD and NECTR-Attn-Tr remain within 0.13 dB across all seven settings.

These results clarify that the predictor and direct training contribute substantially to reconstruction quality. We believe the reconstruction gains come primarily from the new attention backbone and from using the same operator for denoising during training and for regularization during reconstruction.

(C2) Admissible denoisers and PSD. As clarified in (A1), NECTR admits negative eigenvalues. Our intended point was that its normalization preserves PSD when present, although PSD is neither guaranteed for the learned operator nor required for convergence. We will correct this discussion.

At the normalization level, both methods take a symmetric nonnegative matrix KK with positive row sums and produce a symmetric nonnegative stochastic matrix WW. Their maps have the same range but can produce different outputs from the same input KK. Both return W=KW=K when KK is already stochastic. We therefore do not claim a broader admissible matrix class.

(C3) We thank the reviewer for suggesting this optimization. We implemented the fused accumulation for NECTR and repeated the comparison under matched caching policies. The two methods have essentially identical reconstruction costs: NECTR-Attn-Tr/Metro-SSD take 1.620/1.613 s for deblurring, 1.625/1.615 s for SR, and 39.83/39.82 s for MRI. Peak training and cache memory are comparable. We will revise the manuscript acordingly.

(C4) Channel constants and MRI. The required multichannel condition is ker⁡A∩C={0}\ker A\cap\mathcal C=\{0\}, where C\mathcal C is the space of channelwise constant images. For RGB deblurring and SR, both operators act channelwise, so this reduces to Ace≠0A_c e\neq0 for every channel cc. Applying the single-channel theorem to each block gives a full contraction factor max⁡cqc<1\max_c q_c<1. For complex MRI, the proof extends using the Hermitian adjoint and complex Euclidean norm; retaining the real part preserves contractivity. Our response (A2) to Reviewer kY3z gives the detailed argument. We will incorporate these clarifications into the manuscript.


Summary

  1. Added ablations showing how the attention predictor and direct training improve reconstruction quality.
  2. Incorporated the reviewer’s fusion suggestion and established comparable computational cost, while correcting the PSD and pass-count distinctions.
  3. Strengthened the theoretical presentation by explaining the multichannel case and the complex MRI extension.

Thank you for the additional experiments and clarifications. The new ablations indicate that most gains over the original NECTR come from the predictor and direct training. With these choices matched, Metro-SSD and NECTR show similar reconstruction quality and runtime. The explanation of the multichannel condition is also helpful. However, the authors acknowledge that the normalization does not broaden the admissible denoiser class, and I still find the contribution beyond NECTR limited. I will maintain my assessment.

Thank you for considering the additional experiments and clarifications. The feedback helped us strengthen the comparisons with NECTR and clarify the sources of improvement in reconstruction.

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

This paper proposed a denoiser for Plug-and-play image reconstruction with a convergence guarantee. Specifically, the authors propose a symmetric denoiser using Metropolis normalization. With the merits of the proposed algorithm, the convergence of several existing PnP frameworks, such as PnP-HQS and PnP-DRS, can be guaranteed. Experiments on image restoration and MRI reconstruction also demonstrated the competitive performance of the proposed method.

Critical strengths
  1. The mathematical formulation as well as the theoretical analysis are rigorous and clear to follow.
  2. The proposed method is cleaner than the multi-stage processing in NECTR.
Critical weaknesses
  1. This paper is well motivated, and the target is also clear. My biggest concern is that the novelty is relatively limited, since the algorithm is largely based on the NECTR framework. It feels like an improved version of NECTR without the requirement for explicit matrix implementation. I think the author should justify more about their contribution and why this NECTR-based framework is the best design compared to the broad range of other PnP with convergence.
Questions for authors
  1. Theorem 1 is based on the condition that only have a single eigenvalue of 1. During the prediction of the channel aggregation weights for RGB images, is it possible to have multiple eigenvalues of 1, since the channels are processed independently?
  2. Before the reconstruction, the proposed method needs to first obtain an estimation of the reconstructed images in the warm-up stage. Will the performance as well as the convergence heavily depend on this warm-up stage?
3 comments

We thank the reviewer for recognizing the rigor of the analysis and the simplicity of the proposed construction. We address the contribution and the two questions below.

(B1) Contribution and choice of framework.

Our contribution is to demonstrate that hard structural constraints can accommodate expressive denoisers with strong reconstruction performance. Starting from NECTR’s permutation-aggregation framework, we develop an exact operator-level implementation of classical Metropolis normalization that provides a simpler alternative normalization for obtaining stochasticity. We then design a neural weight predictor and train the normalized denoiser directly to achieve strong reconstruction quality.

The resulting method combines global nonexpansiveness and provably contractive PnP iterations after reference freezing with reconstruction quality competitive with, and sometimes better than, GSPnP and DPIR. This combination of flexibility, guarantees, and practical performance motivates our design.

(B2) Multiple unit eigenvalues in RGB images.

Yes. The full RGB denoiser is block diagonal and has a constant eigenvector for each channel. Primitivity applies to each channel block, rather than to the full matrix.

For RGB deblurring and SR, the forward operator also acts channelwise. The single-channel theorem therefore applies to each block provided Ace≠0A_c e\neq0 for every channel cc. Each channel's PnP operator is contractive, and the full operator has contraction factor max⁡cqc<1\max_c q_c<1. Kindly note that, response (A2) provides the detailed argument.

(B3) Dependence on warm-up.

Warm-up improves the reference used to predict aggregation weights; it is a finite initial phase of reconstruction. Its role in reconstruction quality is distinct from its role in convergence.

The convergence guarantee does not require the reference to approximate the ground truth. For any fixed reference satisfying the stated construction assumptions, the subsequent iteration is contractive under the forward-model condition. Warm-up can affect the resulting fixed point and contraction factor, but its success is not a prerequisite for convergence after freezing.

Warm-up does matter empirically. In Table 5, increasing its length from zero to the chosen default improves motion deblurring from 24.82 to 29.23 dB and single-coil MRI from 40.61 to 44.05 dB. Performance then largely stabilizes: increasing warm-up from 30 to 50 iterations changes motion deblurring from 29.23 to 29.22 dB, while increasing it from 150 to 250 iterations changes MRI from 44.05 to 44.11 dB. Thus, warm-up substantially improves quality, while the tested settings show little sensitivity to its precise duration once a suitable reference has been obtained.


Summary

  1. Resolved the multiple-eigenvalue concern through channelwise contraction and its extension to the full RGB operator.
  2. Showed that warm-up substantially improves reconstruction quality, with stable performance across a broad range of tested warm-up lengths. Convergence remains guaranteed after freezing under the stated assumptions.

Thank you for the additional clarifications. After reading the authors rebuttal and comments from other reviewers. My main concern still remains the limited novelty based on the incremental contribution to the NECTR framework, which was also mentioned by other reviewers.

Thank you for considering our rebuttal and the additional clarifications.

Reviewer_kY3zOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Metro-SSD, a trainable linear denoiser that applies classical Metropolis normalization to learned aggregation weights to enforce symmetry and stochasticity. The authors establish convergence guarantees for PnP-HQS and PnP-DRS under a fixed-reference setting and evaluate the method on deblurring, superresolution, and MRI.

Critical strengths

1- The method achieves competitive reconstruction performance, particularly on MRI.

2- The paper includes relevant PnP baselines, implementation details, and warm-up ablations.

Critical weaknesses

1- The main contribution is applying classical Metropolis normalization to the existing NECTR framework, so the novelty in optimization appears limited. The paper does not introduce a new optimization algorithm, and the convergence analysis mainly follows existing results. Moreover, the paper suggests that NECTR requires positive semidefiniteness (PSD), but NECTR's convergence assumptions do not require PSD. Its normalization can also admit negative eigenvalues. It is therefore unclear what theoretical advantage the proposed normalization provides over NECTR.

2- The proof assumes that eigenvalue 1 is simple, whereas the channel-specific aggregation described in the paper produces a block-diagonal operator with multiple constant eigenvectors. Thus, the full multichannel operator is not primitive, and the condition A1≠0A\mathbf{1}\neq 0 is generally insufficient. The authors should establish the appropriate nullspace condition for the multichannel setting and clarify the treatment of complex-valued MRI data.

3- Metro-SSD achieves competitive results but does not consistently outperform existing methods. Since the theoretical contribution is also limited, I do not find the contribution of this paper sufficiently convincing.

Questions for authors

1- What is the main novelty in optimization beyond applying Metropolis normalization and existing convergence results? How does the proposed method differ from NECTR?

2- How does the convergence proof extend to the multichannel implementation, including complex-valued MRI?

3- Can the authors provide fair runtime and memory comparisons against optimized NECTR?

3 comments

(A1) On the contribution to optimization, we agree that we are not proposing a new convergence theory. Our contribution is to show that the hard restrictive assumptions of an existing theory on linear denoisers can accommodate practical denoisers with strong reconstruction performance. We address the common concern that enforcing constraints such as nonexpansivity limits expressiveness and compromises reconstruction quality compared to more flexible nonconvex or heuristic methods. By adapting Metropolis normalization to permutation-based aggregation, we construct denoisers that are linear, symmetric, and stochastic. Despite these global hard structural restrictions, the resulting denoisers achieve reconstruction quality competitive with, and in some settings better than, methods such as GSPnP and DPIR (which are infact very hard to beat even with unconstrained PnP systems). We believe that constructing such denoisers and demonstrating their effectiveness is a meaningful contribution, even though the convergence guarantees build on existing theory.

We agree that Metro-SSD uses the convergence theory for symmetric stochastic denoisers employed in NECTR. We also agree that NECTR neither requires PSD nor excludes negative eigenvalues. Our intended point was that its normalization preserves PSD when present, although PSD is neither guaranteed for the learned operator nor needed for convergence. Metro-SSD directly enforces symmetry and stochasticity without seeking to preserve PSD. We will correct this discussion.

(A2) We thank the reviewer for identifying this concern. The full channel denoiser (acting on R3n\mathbb{R}^{3n}) is indeed block diagonal and need not be primitive. For RGB deblurring and SR, however, both the frozen denoiser and the forward operator act independently on each channel in Rn\mathbb{R}^n. The single-channel theorem therefore applies to each block, provided (Ace≠0A_c e\neq0) for every channel (c)(c). Each channel’s PnP operator is then contractive, and the full operator is contractive with factor (max⁡cqc<1\max_c q_c<1), where (qcq_c) is the contraction factor for channel (c)(c).

The proof over Rn\mathbb R^n extends naturally to Cn\mathbb C^n for complex MRI. We outline the extension below, following the same argument as in the real case.

For z=u+ivz=u+iv, with u,v∈Rnu,v\in\mathbb R^n, the norm is ∥z∥22=∥u∥22+∥v∥22.\Vert z\Vert_2^2=\Vert u\Vert_2^2+\Vert v\Vert_2^2. Since the denoiser WW is a real linear operator, its natural complex extension is W(u+iv)=Wu+iWv.W(u+iv)=Wu+iWv. Its nonexpansiveness over Rn\mathbb R^n therefore gives nonexpansivity in Cn\mathbb C^n: ∥Wz∥22=∥Wu∥22+∥Wv∥22≤∥u∥22+∥v∥22=∥z∥22.\Vert Wz\Vert_2^2=\Vert Wu\Vert_2^2+\Vert Wv\Vert_2^2\leq\Vert u\Vert_2^2+\Vert v\Vert_2^2=\Vert z\Vert_2^2. Moreover, equality holds precisely when both uu and vv are constant vectors.

For the fidelity, we replace the transpose by the Hermitian adjoint, R=(I+ρA∗A)−1.R=(I+\rho A^*A)^{-1}. The matrix A∗AA^*A has an orthonormal eigenbasis with real eigenvalues λj≥0\lambda_j\geq0. The corresponding eigenvalues of RR are (1+ρλj)−1∈(0,1](1+\rho\lambda_j)^{-1}\in(0,1]. Thus RR is nonexp. on Cn\mathbb C^n, with norm preserve precisely on ker⁡(A∗A)=ker⁡A\ker(A^*A)=\ker A.

The remainder follows exactly as in the real case. The linear part of PnP-HQS is Φ=WR\Phi=WR, so ∥Φ∥2≤1.\Vert\Phi\Vert_2\leq1. Suppose equality holds. The unit sphere in Cn\mathbb C^n is compact, so the operator norm is attained at some unit vector vv. Hence 1=∥WRv∥2≤∥Rv∥2≤∥v∥2=1.1=\Vert WRv\Vert_2\leq\Vert Rv\Vert_2\leq\Vert v\Vert_2=1.

Both inequalities must be equalities. By the norm-preservation properties, the first gives Rv∈spanC{e}Rv\in\text{span}_{\mathbb C}\{e\}, while the second gives v∈ker⁡Av\in\ker A. Since Rv=vRv=v on ker⁡A\ker A, it follows that

v∈spanC{e}∩ker⁡A={0},v\in span_{\mathbb C}\{e\}\cap\ker A=\{0\},

where the last equality follows from Ae≠0Ae\neq0. This contradicts ∥v∥2=1\Vert v\Vert_2=1. Thus ∥Φ∥2<1\Vert\Phi\Vert_2<1, showing contractivity for every ρ>0\rho>0.

In our code, we only retain the real part of the complex output, which is again a nonexp. operation. Thus we retain the contractivity. Similar argument is applicable to DRS also.

(A3) Runtime and memory. Append. E.3.2 benchmarks all methods on the same RTX PRO 6000 BW GPU, rebuilding weights/operators during warm-up and caching them after freezing.

Total time (s), ordered as NECTR-Proj (Sinha et al.) / NECTR-Attn / Metro-SSD:

  • Deblurring: 3.16 / 1.76 / 1.59
  • SR: 3.17 / 1.77 / 1.600
  • MRI: 26.41 / 14.48 / 13.70

Metro-SSD ~10% faster than NECTR-Attn for deblurring/SR and ~5% for MRI, with 4–5% faster training steps and comparable training memory. Peak training memory is 54–55 GiB for NECTR-Proj versus 4.54–4.61 GiB for attention variants. Warm-up working memory favors NECTR-Proj: 0.28/0.56 GiB vs 0.90/1.75 GiB (deblurring–SR/MRI).


Summary

  1. Clarified the main contribution: constructing expressive denoisers that combine strong reconstruction quality with exact structural constraints and provable convergence.
  2. Extended the convergence argument to channelwise RGB and complex MRI.

Overall, the rebuttal has addressed several of my technical concerns. I am updating my rating from 1 to 2, but my main concern regarding the limited methodological novelty remains.

Thank you for carefully considering our responses and reassessing the manuscript. The detailed feedback helped us clarify the theoretical assumptions and strengthen the presentation of our contributions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.