acceptodds
Withdrawn from ICLR 2027

-PnP: Plug-and-Play Denoising for Scientific Imaging with Self-Supervised Score Priors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Modern image restoration methods increasingly rely on large-scale supervised training, yet clean targets are often prohibitively difficult to obtain in scientific imaging modalities such as hyperspectral imaging (HSI) and magnetic resonance imaging (MRI). Plug-and-play ADMM (PnP-ADMM) with pretrained diffusion priors offers a promising alternative, but the learned prior distribution can mismatch both the target scientific image distribution and the distribution of ADMM denoiser inputs induced by dual feedback. To mitigate these mismatches, we propose S³-PnP, a self-supervised plug-and-play restoration framework that learns a score prior directly from observation-derived ADMM intermediate states, without clean training targets or externally pretrained weights. Specifically, the learned score corrects the input to the analytical prior update, combining observation-driven corrections with analytical regularization. We further analyze the convergence and reconstruction error of the corrected ADMM algorithm assuming an ideal score and quadratic data fidelity. Experiments on reference-guided HSI restoration, single-image HSI denoising, and MRI denoising demonstrate consistent improvements over competing methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Scientific imaging Denoising, Self-supervised learning, Plug-and-play image restoration, Score matching, ADMM, Test-time learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_zxxJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes S$^3$-PnP, a self-supervised plug-and-play framework for scientific image restoration. Instead of relying on externally pretrained diffusion priors, the method first runs an analytical ADMM solver on the current observation and collects its intermediate denoiser inputs as instance-specific training data. A score model is then learned at test time through self-supervised denoising score matching to characterize the empirical distribution of these ADMM states. During inference, the learned score is used to correct the ADMM intermediate state before applying the original analytical prior operator, thereby combining observation-adaptive learned corrections with explicit model-based regularization. The experimental results show that the proposed method achieves competitive or superior quantitative and qualitative restoration performance compared with various analytical, self-supervised, and diffusion-based methods.

Critical strengths
  1. The proposed method does not rely on an external training dataset or pretrained weights, but instead learns task-specific score corrections from ADMM states generated by the current observation. This design reduces dependence on domain-specific training data and helps alleviate generalization issues caused by distribution shifts between training and test data.
  2. The paper evaluates the proposed method on multiple scientific imaging tasks, including reference-guided hyperspectral image restoration, single-image hyperspectral denoising, and MRI denoising. The diversity of tasks and noise settings provides relatively comprehensive empirical validation of the proposed framework.
Critical weaknesses

The methodological novelty is limited, and the core assumption behind the learned score is not fully convincing. The proposed framework mainly combines an analytical ADMM prior with a test-time learned score correction, while the score itself is obtained through standard self-supervised denoising score matching. More importantly, some ADMM intermediate states may deviate from the ideal solution, contain systematic errors, or lie in undesirable regions of the state space, yet they are still used as training targets. Consequently, the learned network estimates the score of the empirical ADMM-state distribution rather than that of an ideal ADMM-state distribution. This raises concerns about whether the learned correction truly provides a reliable direction for improving reconstruction quality.

Questions for authors
  1. Since some ADMM intermediate states may themselves deviate from the ideal solution, contain systematic errors, or lie in undesirable regions of the state space, these states are nevertheless still used as training targets. Consequently, the learned network estimates the score of the empirical ADMM-state distribution, rather than the score of an ideal-ADMM-state distribution.
  2. A score learned from a single-instance ADMM trajectory may have substantially weaker statistical modeling capability than a large pretrained diffusion prior. It is unclear whether the observed performance gains genuinely come from the proposed task-adaptive score modeling, or are largely attributable to the Deep Image Prior (DIP) effect induced by test-time network fitting.
  3. Since pretrained diffusion-PnP mismatch is a central motivation, the absence of DiffPIR (Zhu et al., Denoising Diffusion Models for Plug-and-Play Image Restoration) or a controlled pretrained diffusion-PnP baseline significantly weakens the experimental support for the main claim.
Reviewer_5UAaOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a plug-and-play ADMM framework for restoring scientific images (hyperspectral and MR images) when neither clean training data nor domain-matched pretrained priors are available. The method first runs a conventional ADMM solver with a hand-crafted prior. It collects the denoiser inputs into a training set, and trains a small network on this set at test time with a generalized denoising score matching objective.

Critical strengths
  1. The paper addresses a practical problem. Scientific imaging often lacks clean data and domain-matched pretrained models.

  2. Reported performance is strong.

Critical weaknesses

Limited applicability and missing cost analysis. Apart from reference-guided HSI restoration, all tasks are denoising. MRI is evaluated only on magnitude-image denoising, although compressed-sensing reconstruction, super-resolution, deblurring, and inpainting are the standard use cases of PnP-ADMM. The test-time cost is not reported against the baselines.

Hyperparameters are selected on the test set using clean references. The reported results appear to depend on hyperparameters chosen on the test data with clean references. Appendix B.5 states that for Dongying the clean high-resolution images are used for parameter selection and evaluation, and that the 25 evaluated M4Raw subjects include the three used for parameter selection. Tables 5-8 show that the solver and correction parameters are set separately for each noise condition and contrast. In a setting motivated by the absence of clean references, this amounts to oracle tuning, and it is unclear whether the baselines were tuned the same way. Could the authors report results with hyperparameters selected on data disjoint from the evaluation set, or fixed across conditions, and describe how the baselines were tuned?

The analytical ADMM baseline is missing. The main tables do not include the analytical ADMM solver itself (\beta = 0), so the contribution of the learned score cannot be separated from that of the analytical solver. Could the authors add a best-tuned analytical ADMM row to Tables 1–3, tuned with the same protocol as S3-PnP?

The distribution of ADMM states shifts once the correction is enabled. The score network is trained on states collected from the analytical solver (the first training set). Once the correction is enabled, the iterates follow a different trajectory and need not resemble those training states. The authors acknowledge this shift for the structure-guided variant, where ADMM is rerun to collect a second training set "to adapt to states generated after the structure correction." The full-space network, however, is trained only on the first training set and then applied at all K iterations. Two observations suggest the correction is not reliable throughout the iterations.

It is unclear what the learned score represents. What is the "clean state" Z_0 for an ADMM denoiser input, and how should \sigma_0 be chosen in practice? Can the authors give intuition, or a diagnostic, for what the correction removes from the states?

The theory does not match the experiments. Theorem 1 assumes an ideal score, a positive-definite quadratic data term, and a firmly nonexpansive analytical map. Can authors clarify all of those assumptions are justified in practice? Moreover, the input Y in the theorem is the perturbed observation itself, so the ADMM-state aspect that motivates the method disappears from the analysis. Also, can authors clarify what does the analysis add beyond standard PnP-ADMM contraction arguments (e.g., Ryu et al., 2019)?

Structure-guided learning is hard to follow and not properly ablated. Section 3.2 mentions column stripes in hyperspectral images without explaining where they come from or why they are a problem for the learned score. The paper also does not explain why orthogonal projectors are needed. The stripe subspace depends on "detected interference support", but the detection procedure is not described.

The writing is hard to follow.

  • "Analytical prior" is never defined in the main text. Only App. B.1 shows that it means hand-crafted regularizers such as local singular-value thresholding, DCT sparsity, and TV/TGV.
  • GDSM is used without being introduced, so it is hard to see what Eq. (3) represents and what minimizing it yields.
  • It is unclear which prior is applied after the first iteration. In fact, the analytical map is applied at every iteration, and the network only modifies its input.
  • Algorithm 1 is not self-contained: several variables are not defined there.
  • The notation is overloaded:
    • calligraphic C denotes the denoiser, while C^k denotes its input;
    • A denotes the image, while H is the forward operator;
    • K is used both as an integer and as an index set; Together with the many hyperparameters set per condition, this makes the algorithm hard to follow and to reproduce.

Based on the above weaknesses regarding the distribution mismatch from the correction, heavy parameter tuning, writing quality, and unsupported theoretical assumptions, I recommend rejection.

Questions for authors

Please refer to the weaknesses section above.

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

The manuscript presents a self-supervised plug-and-play ADMM framework for HSI and MRI restoration. It learns a score prior from ADMM intermediate states derived from the observations, without clean targets or pretrained weights.

Critical strengths

Avoiding clean data is attractive for scientific imaging.

Critical weaknesses

The two mismatches the authors target have both been studied recently. Shoushtari et al. (ICML 2024) analyse prior mismatch and adaptation in PnP-ADMM, and Shrestha et al. (ICLR 2026) address the mismatch between score-based denoisers and the distribution of ADMM iterates. The distinctive element is therefore learning the score without supervision. Yet learning priors from corrupted data alone has also been explored, for example in Noise2Score, GSURE-based training and expectation-maximisation approaches (Rozet et al., NeurIPS 2024). The authors should state precisely how their formulation differs from these lines of work.

More fundamentally, the intermediate states are shaped by measurement noise, reconstruction errors and dual feedback, and their distribution changes during optimisation. A score learned from them may therefore reinforce existing distortions rather than capture clean-image statistics. The authors should explain how this circular dependency is controlled and which noise model the estimation assumes. They should also clarify how much the learned correction contributes compared with the analytical regulariser, and what distribution the score represents when only a single observation is available.

The convergence analysis, as the authors acknowledge, assumes an ideal score. Its relevance to the algorithm actually implemented thus remains uncertain. A sensitivity analysis with respect to errors in the score would be valuable.

Questions for authors

See weakness part.

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

Related papers

Discussion (0)

Sign in to comment.