acceptodds
Under review as a conference paper at ICLR 2027

Reconstruction-Aware Distribution Matching for Inverse Problems with Unpaired Data

Abstract

Recovering the forward operator of an imaging system from unpaired data avoids both calibration hardware and the clean/degraded pairs that supervision requires—pairs that, for a real lens, are often impossible to acquire. Existing unpaired approaches that explicitly estimate the forward operator by distribution matching evaluate a candidate only through the measurements it generates, whose distribution should match that of the real ones. Components suppressed by the operator are barely present in the degraded images, and therefore barely constrain the operator under such a criterion. Yet inversion precisely tries to recover them. As a result, two operators that are nearly indistinguishable as forward models can invert very differently. We address this by adding a comparison on the clean side: real degraded images, once restored, should be distributed like clean ones. Restorations are computed by a differentiable plug-and-play algorithm parameterized by the operator being learned. On spatially varying PSF calibration and blind super-resolution, the method recovers more accurate operators than degraded-side matching alone, and closes much of the gap to restoration with the true operator.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
inverse problems

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_2k4VOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This article proposes a "reconstructed perception" distribution matching method to address the challenge of estimating unknown forward operators (such as blur kernels, PSF fields, etc.) from unpaired data in unsupervised image inverse problems. The existing methods based on pure degraded end distribution matching often find it difficult to constrain high-frequency or low signal-to-noise ratio information that is strongly suppressed by operators, resulting in operators that appear similar in the forward model performing differently in inverse operations (reconstruction). Therefore, based on the original degraded end matching (measure matching), this article innovatively introduces the reconstruction matching loss of the clean end. Real time reconstruction of real degraded images using differentiable alternating direction multipliers/PnP solvers (such as DPIR), and requiring the reconstructed image distribution to approximate a clean image distribution. Using conditional flow matching (CFM) combined with ensemble KL (IKL) divergence to measure distribution differences; Addressing the issue of high memory consumption in iterative solvers through truncated backpropagation; We have designed an adaptive gradient modulus balance heuristic strategy to prevent the reconstruction gradient from dominating the optimization process due to iterative amplification. In tasks such as uniform blurring, spatial variation PSF calibration, blind super-resolution, and real camera lens calibration, this method is significantly superior to traditional only degraded end matching methods, significantly narrowing the gap with the reconstruction effect of real operators

Critical strengths
  1. Deep theoretical insight and motivation (Theorem 1): The paper rigorously proves theoretically that pure degraded end matching attenuates information extremely quickly on low signal-to-noise ratio components ($\ Theta (\ rho ^ 2) ),whilereconstructionerrorattenuatesitssensitivitymuchslower(), while reconstruction error attenuates its sensitivity much slower (\ Theta (\ rho) $), accurately hitting the fundamental pain point of existing non paired operator estimation methods, and providing strong theoretical support.

  2. Clever end-to-end differentiable reconstruction adaptation: By embedding the classic PnP IR algorithm (DPIR) into the training loop and combining truncated backpropagation with dynamic gradient normalization, the expensive iterative differentiation of VRAM and multi loss scale imbalance problems have been successfully solved. The engineering implementation is elegant and has strong universality.

Critical weaknesses

1.Although the author discussed the differences between biased loss and unbiased loss in the appendix, the overall framework still heavily relies on fixed pre trained denoising (prior in DPIR). If the pre trained denoising device itself has inherent artifacts or biases, it may mislead the operator estimation.

  1. Due to the need to run an iterative solver (DPIR) with an internal deep neural network in each iteration for the reconstruction branch, its optimization time significantly increases (about 5.7 to 6.6 times that of pure degraded end matching), which limits its scalability in ultra large models or ultra high definition image scenes.

3.In practical operation, the forward noise and the required noise level ($\ sigma $) of the denoising device are treated as fixed hyperparameter inputs. However, in complex real-world blind image restoration scenarios, the noise is often unknown and spatially variable, and the robustness evaluation of the method for noise parameter misestimation is not sufficient.

  1. In order to save video memory, the author adopted a truncation strategy of backpropagation through only the last $K0 {bp} iteration(iteration (K0 {bp}=1 $in the experiment). This strong truncation may not fully capture long-term dependencies in multi-step iterative solvers, weakening the theoretical convergence guarantee.

5.In the core practical application of "real shot calibration" (Section 5.4), due to the lack of Ground Truth, relying only on visual contrast and subjective qualitative analysis, there is a lack of quantitative indicators to support it in real and complex scenes.

Questions for authors
  1. Regarding solver steps and truncation: In the experiment, the total number of iterations for DPIR was $K=8 $, while the number of return steps was $K {bp}=1 $. Has the author tested the extent of improvement in operator estimation accuracy and the boundary of video memory overhead by increasing $K_ {bp} $?

  2. Regarding the sensitivity of noise hyperparameters: In real-world tasks, if the preset denoising level $\ sigma $deviates from the actual data (e.g. by 50%), to what extent will the convergence stability of the algorithm and the final estimated operator degrade?

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

This paper estimates an unknown forward operator from unpaired clean and degraded images. The operator is learned by matching the distribution of generated measurements to the degraded-image distribution and the distribution of DPIR reconstructions to the clean-image distribution, using IKL-based gradients.

The main novelty lies in incorporating reconstruction-side distribution matching into operator learning through a fixed reconstruction algorithm parameterized by the same candidate operator. The paper also presents a sensitivity analysis in the low-SNR regime and evaluates the method on deblurring, spatially varying PSFs, and blind super-resolution.

Critical strengths

S1. The practical performance gains are impressive. In particular, the method substantially outperforms Restormer and OmniLens in reconstruction quality in the measured spatially varying chromatic PSF setting in Table 1.

S2. The paper quantifies how the sensitivity of measurement matching to relative operator errors decays faster at low SNR than that of reconstruction risk. This analysis is useful for understanding the limitations of existing training objectives.

Critical weaknesses

W1. The main contribution is the addition of a reconstruction-side distributional loss to the existing method in [1]. The main idea resembles cycle consistency and is not itself new, as related approaches have already been explored in inverse problems [2, 3]. The novel aspect of this paper is its design that connects the two losses through the same operator without training a separate inverse network. However, since the key tools used to implement this approach, IKL [1] and DPIR [4], are also existing methods, I consider the methodological advance to be limited.

W2. The sensitivity analysis explains the limitations of measurement matching, but the fact that the reconstruction-side loss is more sensitive does not necessarily mean it is a better objective for operator estimation. The unbiased-variant experiment in Appendix E.7 partially alleviates concerns about bias, but does not resolve this logical gap. The paper should clarify to what extent the sensitivity analysis supports the validity of the proposed training objective.

Questions for authors

Q1. Meanti et al. [1] also compared their method against single-image blind kernel estimation methods. Could the authors provide similar comparisons under applicable uniform-blur settings?

Q2. Would it be possible to conduct experiments with a longer backpropagation horizon? I would be interested in seeing how kernel error, reconstruction performance, and computational cost change compared to the current setting of K_bp=1.

Q3. The clean-image prior and the characteristics of the solver used during training appear to play an important role in the performance gains. Further analysis of their influence would be valuable. For example, it would be useful to examine whether the improvements persist when the learned operator is applied to a different non-blind solver.

[1] Meanti et al., Unsupervised Imaging Inverse Problems with Diffusion Distribution Matching, ICCV 2025.

[2] Lim et al., CycleGAN with a Blur Kernel for Deconvolution Microscopy: Optimal Transport Geometry, IEEE TCI 2020.

[3] Sanghvi et al., Photon-Limited Blind Deconvolution Using Unsupervised Iterative Kernel Estimation, IEEE TCI 2022.

[4] Zhang et al., Plug-and-Play Image Restoration with Deep Denoiser Prior, IEEE TPAMI 2021.

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

The paper proposes a reconstruction-aware distribution matching framework for estimating the forward operator in imaging inverse problems with unpaired clean and degraded data. The paper is built on the Unsupervised Imaging Inverse Problems with Diffusion Distribution Matching of Meanti et al. (2025). Authors compare the images reconstructed from real measurements follow a distribution similar to that of the clean images. The proposed method combines measurement-side and reconstruction-side distribution matching, using a differentiable plug-and-play reconstruction algorithm (DPIR) and flow matching-based gradient estimation. The method is evaluated on spatially invariant deblurring, spatially varying PSF calibration, blind super-resolution, and real-lens calibration.

Critical strengths

[1] The paper addresses an important limitation of measurement-only distribution matching by incorporating reconstruction quality into the operator estimation. The proposed method solves an unaddressed challenge in Meanti et al. (2025), addressing the difficulty of distinguishing operators that generate similar measurement distributions but behave differently during reconstruction.

[2] The proposed method improves operator estimation and reconstruction quality across different imaging inverse problems. The improvement in spatially varying PSF calibration is remarkable.

Critical weaknesses

[1] The theoretical analysis uses a simplified setting with MMSE reconstruction, while the actual implementation uses DPIR and IKL. Although the authors acknowledge this difference, a clearer explanation of how the theoretical findings relate to the practical method would help strengthen the connection between the theory and the experimental results.

[2] The method involves important training choices, like initialization, truncated backpropagation, and gradient balancing. It would be helpful to better understand how sensitive the results are to these choices and whether the improvements remain consistent under different settings.

[3] The quality of Figure 2 is not very good; please improve the quality.

Questions for authors

Could the authors provide further justification regarding how the theoretical sensitivity results relate to the practical objective based on IKL and the DPIR reconstruction algorithm?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.