acceptodds
Under review as a conference paper at ICLR 2027

PriorShare: Benchmarking and Understanding Hallucination in Diffusion-Based Scientific Imaging

Abstract

In scientific imaging, reference images are often unavailable, so a small measurement residual is the main check on a reconstruction from a reusable diffusion prior. The residual is insensitive to errors in source components that the measurement leaves unresolved, which we study as hallucinations. We introduce the prior share, the fraction of source energy in the directions a given operator and noise level leave unobservable. It is computed from the operator and representative source data without a test reference, and unlike operator-only summaries it accounts for how source energy is distributed across unresolved directions. We build the PriorShare benchmark around three new settings: weak-lensing mass mapping, galaxy PSF deconvolution, and reconstruction from real wide-field fluorescence observations. With multi-coil MRI and black-hole imaging added, it evaluates nine diffusion-prior solvers. In a controlled study that varies only the operator, agreement between residual and PSNR rankings decreases as the share grows. At a share of 72%, one solver reaches a relative residual of while of its squared reconstruction error lies in the discarded directions, mostly in the unmeasured mean level. In the scientific settings, four of nine solvers on BioSR and five on Galaxy incur larger squared error in the unresolved components than zero filling. On noiseless galaxy deconvolution, an untuned inverse without a learned prior outperforms all nine solvers in PSNR and SSIM. PriorShare thus reports, beside data fit and reference fidelity, how much source energy a residual cannot check.

Primary area:
datasets and benchmarks
Keywords:
scientific inverse problems, diffusion priors, inverse-problem solvers, benchmark evaluation, computational imaging

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

The paper studies the discrepancy between measurement consistency and source recovery in diffusion-based scientific imaging. It defines an unresolved subspace comprising exact null-space directions and weakly observed directions whose source-scale signal falls below the noise level, and introduces prior share as the fraction of source energy contained in this subspace. The authors construct a benchmark spanning weak-lensing mass mapping, Galaxy PSF deconvolution, real wide-field fluorescence reconstruction, MRI, and black-hole imaging, and use it to compare nine diffusion-prior solvers with classical reconstruction baselines.

Critical strengths

[1]It provides a useful way to distinguish good data fit from reliable recovery of weakly constrained components.

[2]The scientific benchmark construction and classical baselines are strong empirical contributions.

Critical weaknesses

[1]The paper does not introduce a new reconstruction method, nor does it derive a new benchmark construction principle from the proposed analysis. The two propositions largely formalize standard consequences of linear inverse problems: multiple reconstructions can be measurement-consistent due to unobservable directions, and a small measurement residual does not in general imply small reconstruction error. These conclusions follow directly from the singular-value decomposition and are closely related to classical null-space/ill-posedness arguments. The paper reframes such ambiguity under the notion of “hallucination,” but it is not clear that this framing yields substantially new theory beyond existing null-space or range-space interpretations. The main novelty therefore appears to lie in the diagnostic metric and benchmark design rather than in reconstruction methodology or inverse-problem theory.

[2]The definition of prior share appears basis-dependent in the presence of repeated singular values. Eq. (1) defines the unobservable set through individual singular directions viv_i and their source energies SiS_i. However, when a singular value has multiplicity greater than one, the corresponding singular-vector basis is not unique. As written, the membership of individual directions in UU, and hence the resulting prior share, can depend on the arbitrary choice of basis within a degenerate singular subspace. In the reported experiments, this ambiguity is largely avoided by using a fixed Fourier basis for operators that are diagonal in that basis, but this suggests that the basis is effectively part of the definition in practice. A basis-invariant formulation would make the concept more generally applicable and theoretically cleaner. [3]The benchmark evidence is only partially representative of real inverse problems, and some comparisons appear insufficiently stable.

Questions for authors

Please refer to weaknesses.

1.The two propositions seem to follow directly from standard SVD-based properties of ill-posed linear inverse problems, namely non-uniqueness in unobservable directions and the fact that measurement residual only constrains error in sufficiently observable directions. What is the precise theoretical contribution beyond existing null-space/range-space analyses of inverse problems and hallucination? Is the intended novelty primarily the prior-share diagnostic rather than the propositions themselves?

2.Do the propositions or prior share lead to any new reconstruction rule, regularization strategy, solver adaptation, or risk guarantee? If not, should the contribution be interpreted primarily as a diagnostic and benchmarking framework?

3.While the paper documents how prior share is computed for each benchmark setting, it does not provide a systematic protocol for constructing benchmark instances with prescribed prior-share levels. Can you discuss whether such a construction protocol is feasible, and whether the current controlled suite can be extended toward it?

4.To what extent do the main conclusions persist on genuinely acquired inverse-problem measurements rather than simulated or synthetically degraded observations? In particular, could the authors clarify which results are driven by real measurements versus controlled synthetic forward models, and whether the observed residual–reconstruction mismatch is equally strong in the real-observation settings?

4 comments
AuthorsOpenReview ↗

Response to Reviewer wbgo (Part 1/2)

Thank you for your detailed review and constructive questions. Below we address each point in detail.

W1 and Q1 · Scope of the theoretical claims

Extension beyond classical null-space analysis. In Section 2, we clarify that the two statements build on the exact null-space analysis of Bhadra et al. and extend it to a broader unobservable set containing both exact null-space modes and directions drowned in measurement noise. By combining the forward operator with the source energy spectrum, our framework quantifies how much target energy lies beyond measurement verification and provides clear error bounds, reconstruction corrections, and solver selection rules.

W2 · Basis invariance under repeated singular values

Basis-invariant formulation. For translation-invariant image populations, the Fourier basis used in our experiments already diagonalizes both the forward operator and the source second moment. To make this basis invariance explicit in general, we clarified Eq. 1 in Section 2 by specifying that within any repeated-singular-value subspace, the singular directions are aligned with the eigenvectors of the restricted source second moment and the directional energies are their eigenvalues. This specification ensures that both the unobservable set and the prior share are strictly independent of the choice of basis, and Section 2 also provides the corresponding basis-free linear MMSE error fraction. The table below compares the Fourier and empirical eigenbasis shares in percent across all tasks with repeated singular values, from Table 19 in Appendix I.5.

SettingSamples per dimensionHard: FourierHard: eigenSoft: FourierSoft: eigen
Galaxy 0.8″ / 1.18″ / 1.8″> 2401.93 / 3.44 / 5.491.74 / 3.40 / 5.47
BioSR> 907.867.86
Weak lensing, noise 0.01, 32×3216018.0718.4924.4823.83
Weak lensing, noise 0.04, 32×3216050.8248.1952.9050.71

Comparison across tasks. On Galaxy, BioSR, and 32×32 weak-lensing patches, where training samples outnumber the subspace dimension, the empirical eigenbasis and Fourier shares agree closely, differing by at most 0.2% on Galaxy and BioSR. On full-size weak lensing, estimating a 16,383-dimensional degenerate eigenbasis from 10,000 patches causes a finite-sample drift from 7.98% to 4.68%, which we verify by subsampling at fixed patch size in Appendix I.5. Because both bases coincide analytically for a statistically homogeneous field on a periodic domain, we report the Fourier value.

Q2 · Risk guarantees, reconstruction rules, and solver selection

Error bound and risk threshold. Outside the unobservable set, Remark 2 shows that the measurement residual directly bounds the reconstruction error. Inside the unobservable set, where measurements provide no constraint, the prior share equals the expected squared error of a simple zero-fill prediction. This zero-fill error sets a clear risk threshold in Table 6, since any solver whose error in the unobservable set exceeds zero fill is adding more hallucination than signal.

Reconstruction with the training mean. This decomposition also gives a direct check on the unverified part of a reconstruction. Replacing a solver's component in the unobservable set with that of the training-set mean image leaves the measurement residual unchanged. On operator O1, the pseudo-inverse filled with the training mean reaches 29.1 dB, above seven of the nine diffusion solvers, and the same replacement lifts RED-diff from 11.5 dB to 29.0 dB, with about 9 dB of the gain coming from the 509 frequencies beyond the mean level, as shown in Appendix I.4. The replacement lowers PSNR for six of the nine solvers, so whether the training mean improves a solver's unverified components depends on the solver, as Section 5.3 reports.

Measurement design and solver selection. Before data collection, minimizing the prior share guides measurement design. Adding the image mean alone reduces the prior share of O1 from 72% to 5.6%, and targeting a 10% share on weak lensing sets the required survey noise level to 0.011. During reconstruction, comparing each solver against the zero-fill threshold in Table 6 guides regularization and solver choice, showing that RED-diff and DPnP stay below the zero fill in all three settings of Table 6, while DiffStateGrad, the solver with the smallest residual on real BioSR data, exceeds it on every image.

AuthorsOpenReview ↗

Response to Reviewer wbgo (Part 2/2)

Step-by-step construction protocol. Appendix I.9 defines a three-step procedure to construct benchmark tasks at a target prior share before any reconstruction. First, for a given training set and a family of forward operators with a tunable parameter such as noise level, PSF width, or frequency mask, we compute the training image energy along each singular direction once. Second, because the unobservable energy grows monotonically as noise increases or as frequencies are masked, we use bisection on a continuous parameter or cumulative summation on discrete frequencies until the unobservable energy fraction matches the target share. Third, we fix that parameter value and generate test observations from held-out source images.

Extending the controlled suite. Applying this bisection procedure to the weak-lensing noise level yields noise standard deviations of 0.011, 0.018, 0.029, 0.042, and 0.059 for target shares of 10%, 20%, 30%, 40%, and 50%. In our controlled study, sorting the 510 unmeasured frequencies of operator O1 by training energy and unmasking them step by step creates a discrete benchmark ladder spanning 0.05% to 5.6% when the mean is measured and 67% to 72% when it is not, as detailed in Appendices I.4 and I.9.

W3 and Q4 · Evidence on real measurements and benchmark stability

Real versus simulated observations. Appendix I.6 explicitly documents the source of the observations in each benchmark task, as summarized in the table below.

SettingObservation source
BioSRReal wide-field observations acquired by a physical microscope
MRIAcquired multi-coil k-space data with retrospective undersampling
GalaxyReal astronomical survey images with a simulated PSF blur
Weak lensing, black-hole imaging, controlled studyFully simulated forward measurements

Mismatch on real BioSR observations. On the 50 real wide-field microscope observations in BioSR, we compare each solver's relative measurement residual, PSNR, and error ratio to zero fill in the unobservable set against the SIM reference in the table below, from Tables 20 and 21 and Fig. 6. Every solver fits the wide-field data more tightly than the 16.2% SIM reference residual, yet a smaller residual still does not mean higher quality. DiffStateGrad, denoted DSG in the table, has the smallest residual at 1.75% but the worst PSNR at 21.6 dB and 2.00 times the zero-fill error in the unobservable set, because pushing the residual below the reference floor set by noise and model mismatch only overfits unmeasured frequencies. The corresponding figures can be seen here.

MetricDSGDPSRED-diffDMPlugDPnPDAPSABMSDMILOSITCOMSIM reference
Residual (%)1.754.024.526.517.137.928.509.409.9716.2
PSNR (dB)21.6024.9226.5623.8724.8324.5023.4725.9323.63
Unobservable error / zero fill2.000.710.401.040.670.761.000.440.98

Stability of the benchmark comparisons. Re-selecting hyperparameters on multiple validation images for weak lensing, Galaxy, and BioSR and evaluating 200 unfiltered weak-lensing patches confirm our conclusions, as shown in Appendices I.1 to I.3. Seven of the nine weak-lensing solvers keep the exact same test PSNR, and the physical-error ranking on the 200 unfiltered patches matches the revised Table 3. Only SITCOM shifts noticeably by 1.8 dB on weak lensing and 0.9 dB on BioSR, while the top and bottom ranks and comparisons to classical baselines remain unchanged.

We appreciate your time and valuable feedback. We hope our responses address your concerns, and we would be grateful if you could reconsider your score.

Dear Reviewer wbgo,

Thank you again for your thoughtful review. We’ve posted a reply to your questions about the basis invariance of prior share and the evaluation on real measurements. When you have a chance, we’d be grateful if you could take a look. If you have any further questions, we’d be happy to discuss them. If our reply addresses your concerns, we’d appreciate it if you could reconsider your rating.

Thank you for the detailed responses and additional analyses. I would like to emphasize that, in my view, developing a meaningful metric for evaluating data and establishing a systematic protocol for constructing benchmark data that meet specified metric levels are equally important aspects of benchmark design. Moreover, I noted that Q3 was not explicitly addressed as a separate item in the response. I see that the additional analysis in Appendix I.9 adequately addresses this concern. I will raise my score.

Reviewer_7uTyOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PriorShare, a benchmark and diagnostic framework for diffusion-based scientific imaging. It estimates the fraction of source energy lying in directions that measurements cannot reliably constrain, using the acquisition operator, noise assumptions, and source statistics. The benchmark evaluates nine diffusion-based solvers across five scientific imaging tasks, including three newly introduced settings. Controlled experiments and error decomposition illustrate why measurement consistency alone is insufficient to assess reconstruction fidelity.

Critical strengths
  1. The benchmark provides useful scientific coverage and practical comparisons. The addition of weak lensing, galaxy deconvolution, and real BioSR observations extends evaluation beyond existing settings. Using a shared prior within each task, domain-specific classical baselines, and scientific metrics makes the comparisons informative for practical applications.

  2. The diagnostic analysis offers insight beyond aggregate reconstruction scores. Weighting poorly observed directions by source energy helps explain why losing a small number of measurement directions can still matter substantially. The matched operator control, training-mean baseline, and error decomposition provide concrete evidence of discrepancies between measurement consistency and reconstruction accuracy. The authors also explicitly acknowledge important limitations of these analyses.

Critical weaknesses
  1. The evaluation protocol leaves substantial uncertainty about the robustness of the benchmark conclusions. Several settings, including weak lensing and Galaxy, use only one validation image for hyperparameter selection. Furthermore, the weak-lensing test subset was revised after observing poor DPS results on near-flat patches (Appendices C.4 and E.1). Although these choices are disclosed, they may influence method rankings and the apparent prevalence of failure cases. More robust validation and evaluation on the complete held-out distribution are needed to establish the reliability of the reported comparisons.

  2. The additional explanatory value of prior share remains insufficiently established. In the main high-share example, approximately 67 percentage points of the reported 72.2% share arise from the DC component. Although the authors investigate this effect, the evidence for broader applicability to structural reconstruction errors remains limited (Section 5.3). Moreover, the Galaxy share uses an assumed 8-bit precision floor despite the absence of added measurement noise (Appendix B.2). Consequently, its interpretation depends on an analysis convention whose connection to the implemented acquisition process requires stronger justification.

Questions for authors
  1. valuation robustness—this is the main issue affecting my assessment: Do the principal method rankings and failure conclusions remain stable with a larger validation set and with near-flat weak-lensing patches retained? Please provide supporting results or clarify what existing evidence establishes this robustness.

  2. Interpretation of prior share: What evidence demonstrates its additional explanatory value after separating mean/DC energy from structural variation? Please also clarify the justification for treating source-image quantization as an observation-noise floor in the noiseless Galaxy experiment.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 7uTy (Part 1/2)

Thank you for your thoughtful review and constructive questions. Below we address each point, with full details in Appendix I of the revised manuscript.

W1 and Q1a · Multi-image hyperparameter selection

Weak-lensing selection. We re-selected all nine solvers on 16 validation patches under four rules fixed in advance. Single uses PSNR on the original patch, Multi the mean PSNR over the 16 patches, Multi-MSE their mean error, and Multi-sig the mean PSNR over the 12 non-flat patches, which the revised Table 3 reports. The table below gives test PSNR on the 50 patches, with full details in Table 12. Seven of the nine solvers reach the same test PSNR under Single and Multi-sig, and the three multi-image rules yield nearly identical rankings. RED-diff and ABMS stay at the top, DMILO stays last, most diffusion solvers remain ahead of Kaiser-Squires, and Multi-sig mainly refines SITCOM by 1.8 dB.

MethodSingleMultiMulti-MSEMulti-sig
RED-diff30.2930.2930.2930.29
ABMS30.1930.1930.1930.19
DPnP29.4628.8829.4629.46
SITCOM27.0128.8228.8228.82
DAPS28.7528.2028.7528.75
DMPlug28.6628.5228.5228.66
DPS28.1028.1028.1028.10
DiffStateGrad27.2627.2627.2627.26
DMILO24.3725.9425.4724.97
Kaiser-Squires27.59

Galaxy and BioSR selection. On Galaxy, six of the nine solvers keep their original configuration under five-image selection, and the top five and bottom ranks remain unchanged. On BioSR, re-selecting on 16 validation pairs keeps RED-diff first and DiffStateGrad last, while SITCOM rises from seventh to fifth by 0.9 dB and the remaining solvers change by at most 0.12 dB. Full results are in Appendices I.2 and I.3 and Tables 15 and 16.

W1 and Q1b · Retaining near-flat weak-lensing patches

Unfiltered test evaluation. We evaluated the convergence RMSE, the error in physical units, on 200 test patches drawn without the flatness filter, 47 of them near-flat, as shown in the table below and Table 14. In physical units every solver has a smaller error on near-flat patches than on the others, and the ranking over all 200 patches matches the revised Table 3. Evaluating four representative solvers across all 2,544 held-out patches gives virtually identical errors and the same order.

MethodAll (200)Near-flat (47)Others (153)
RED-diff0.00990.00380.0118
ABMS0.01010.00320.0122
DPnP0.01090.00430.0130
SITCOM0.01140.00350.0138
DAPS0.01170.00710.0131
DMPlug0.01200.00860.0130
DPS0.01240.00820.0136
Kaiser-Squires0.01320.01010.0141
DiffStateGrad0.01350.01090.0144
DMILO0.01900.01710.0195

Why near-flat patches were excluded. Near-flat patches contain almost no lensing signal, and because PSNR normalizes by each patch's dynamic range, it stops being comparable once that range approaches zero. On those 47 near-flat patches, predicting an all-zero image gives a physical error of 0.0016, lower than every solver. Errors on near-flat patches therefore measure spurious structure added to empty background rather than signal recovery, as detailed in Appendix I.1.

AuthorsOpenReview ↗

Response to Reviewer 7uTy (Part 2/2)

Tasks that already measure the mean. In tasks such as BioSR microscopy and Galaxy deconvolution, the forward operator directly preserves the image mean, so the unobservable set consists purely of high-frequency structural details. Even when global brightness has no ambiguity at all, several solvers still produce larger structural errors in unmeasured frequencies than a simple zero fill, as shown in Table 6.

Controlled experiment without mean ambiguity. Modifying operator O1 to measure the image mean while leaving high-frequency diagonal modes unmeasured lowers the prior share from 72% to 5.6%. Even with the mean known, a small data residual can still hide severe structural errors: DAPS has the third-smallest residual yet the second-worst PSNR, with 95% of its error in those unmeasured structural modes. On O1 itself, fixing the mean of RED-diff raises its PSNR to 20 dB, and also fixing those unmeasured structural modes adds nearly 9 dB more, as shown in Appendix I.4. The corresponding figures can be seen here.

W2 and Q2b · The 8-bit floor in Galaxy

Physical reason for the 8-bit floor. The Galaxy sources are 8-bit survey images, so the 3.4% share treats one 8-bit step as the level below which a blurred frequency can no longer be told apart from zero. In exact floating-point arithmetic the blur has no zero frequencies and the share would be 0, and the 3.4% is the value for an acquisition recorded at the precision of the source data, as stated in Appendix I.8.

Verification with explicit 8-bit noise. Adding 8-bit quantization noise directly to the Galaxy observations changes eight of the nine diffusion solvers by at most 0.15 dB, and all five solvers that exceeded the zero fill in Table 6 continue to do so. Only the two methods that inverted fine detail near that limit drop: SITCOM by 1.3 dB across the zero-fill threshold and unregularized conjugate gradients by 1.1 dB, confirming that the 8-bit floor marks where noiseless inversion breaks down. Full results are in Table 23 and Appendix I.8.

We hope these results and clarifications address your concerns, and we would be glad to discuss any point further.

Dear Reviewer 7uTy,

Thank you again for your thoughtful review. We’ve posted a reply to your questions about the evaluation protocol, including hyperparameter selection and the treatment of near-flat weak-lensing patches. When you have a chance, we’d be grateful if you could take a look. If you have any further questions, we’d be happy to discuss them.

Thank you for the detailed responses and additional experiments. The authors have adequately addressed my main concerns regarding evaluation robustness, the inclusion of near-flat patches, and the interpretation of prior share. The additional analyses provide sufficient evidence to support the reliability of the conclusions. I am satisfied with the clarifications and will maintain my original rating of 3 (Weak Accept).

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

In the context of scientific imaging with linear operators, this paper proposes PriorShare to quantify the fraction of source energy that are underdetermined given the measurement. The experimental results over several scientific imaging problems show that 1) better data fit does not indicate better reconstruction and 2) many existing solvers fail to transfer to different operators with fixed configuration.

Critical strengths

The benchmark problems are diverse and follow useful scientific settings.

Critical weaknesses
  • I fail to see the practical value of the proposed PriorShare. Based on Sec.5.2, the paper intended to propose it as a indicator of how well the residual correlates with PSNR. But the evidence looks weak as the PSNR also depends on the solver algorithm, and the PriorShare depends on the distribution it's computed over.
  • The paper is poorly written and hard to parse.
    • Many notations are used without definition. for example, the operator PU(A)P_{U(A)}, the distribution pp.
    • The whole experiment section (section 5) do not explain the key message behind the experimental results. Instead, the paper mostly repeats the numerical differences that are already present in the table or figure.
Questions for authors
  • The proposed PriorShare heavily depends on the source distribution through representative source data. In practice, how do we define source distribution and how to get representative source data?
  • A simple mean shift in the null-space can cause big inflation to PriorShare because of the second moment term. More generally, how sensitive is PriorShare to the mismatch between the estimated source distribution and the actual groundtruth?
5 comments
AuthorsOpenReview ↗

Response to reviewer hgSa (Part 1/2)

Thank you for your detailed review and constructive feedback. We address each of your points below.

W1 · Practical value of the prior share and Section 5.2

What the prior share measures. We believe there may be a misunderstanding here, and we have clarified its role in Section 2. Computed before any reconstruction from the operator, noise level, and domain source samples, the prior share measures the fraction of source energy in unresolved directions that data residuals cannot verify. Across our benchmark settings in revised Table 7 below, this unverified share ranges from 3.4% to 72%.

OperatorControlled operator O1Galaxy blurBioSRWeak lensing, noise 0.01Weak lensing, noise 0.04
Prior share72%3.4%7.9%8%39%

How it is used in practice. Before data acquisition, this share warns when residuals alone cannot certify a reconstruction and ranks measurement upgrades: on O1, adding the image mean cuts the unverified share from 72% to 5.6%, and in weak lensing a noise threshold of 0.011 keeps it within 10% in Appendix I.9. Where references exist, comparing each solver's unresolved error against a zero fill in Table 6 separates detail recovery from hallucination.

Real BioSR observations. On the 50 real wide-field microscopy pairs of BioSR, all nine solvers fit the measurements more closely than the SIM reference at 16.2% residual. Even so, DiffStateGrad has the smallest residual at 1.75% yet the lowest PSNR at 21.6 dB and twice the zero-fill error in unresolved frequencies, whereas RED-diff and DMILO stay below half in revised Section 5.4. The corresponding figures can be seen here.

Solver dependence in Section 5.2. While a rank correlation across nine solvers varies with the solver set, the conclusion of Section 5.2 rests on the matched pair O1 and O1R, which differ only in whether 510 frequencies are measured. Measuring them raises the agreement between residual and PSNR rankings from 0.4 to 1.0, and O1R stays above O1 in all nine leave-one-solver-out subsets in Appendix I.4. Tables 2 and 18 also report each solver's own error share in those unmeasured frequencies, which is independent of the other solvers. The corresponding figures can be seen here.

AuthorsOpenReview ↗

Response to reviewer hgSa (Part 2/2)

W2 · Notation and key messages of Section 5

Notation. We have added explicit definitions at first appearance in Section 2 for the target source distribution pp, the orthogonal projector PU(A)P_{\mathcal U(\mathcal A)} onto the unresolved subspace, and the singular system sis_i, uiu_i, viv_i, and we checked the entire manuscript to ensure every mathematical symbol is defined when first introduced.

Key messages in Section 5. We also refined Section 5 so each subsection states its main experimental takeaway at the outset. ections 5.1 to 5.4 show that solver rankings on one operator do not transfer to another, that residuals reflect reconstruction quality only when high-energy frequencies are measured, that unmeasured-frequency errors dominate O1 where a simple training-mean fill reaches 29.1 dB and beats seven of the nine solvers, and that comparing unresolved errors against a zero fill detects hallucinated detail across the scientific tasks.

Q1 · Defining and obtaining representative source data

Defining and obtaining source data. In practice, the source distribution is the target population of objects to be imaged, expressed in the same units and preprocessing seen by the solver. One can estimate the share directly from the prior's training set or, when reusing a released checkpoint, a small held-out validation set from the same domain, as we do in Appendix B.2 for weak lensing, Galaxy, BioSR, and MRI without ever using test ground truth.

When source samples are uncertain. Even without source samples, the operator and noise level still determine the unresolved subspace U\mathcal U, as source data only weight its energy fraction, while computing the share across several plausible populations gives a reference range before reconstruction.

Q2 · Sensitivity to source-distribution mismatch

Null-space mean shift. With source mean μ\mu and covariance Σ\Sigma, E∥PUx∥2=∥PUμ∥2+tr(PUΣ)\mathbb E\|P_{\mathcal U}x\|^2=\|P_{\mathcal U}\mu\|^2+\mathrm{tr}(P_{\mathcal U}\Sigma), so a mean shift in U\mathcal U directly changes the numerator, as you noted. In the controlled setting, 67% out of the 72% share comes from the unmeasured global brightness, and we include this term because solvers miss it too, with 84% of RED-diff's O1 error in brightness alone.

Sensitivity across splits and calibrations. As summarized in the table below from revised Table 7 and Appendix B.3, disjoint training halves differ by at most 0.13 percentage points, and switching to held-out test sources or refitting the BioSR acquisition model shifts the share by less than 1.6 percentage points, which is small against the spread from 3.4% to 72% across tasks.

PerturbationShare
Brightness of the controlled-study sources, -20% / 0 / +20%63.7% / 72.2% / 78.4%
Weak lensing, training vs. test sources7.98% vs. 9.53%
Galaxy, training vs. test sources3.44% vs. 3.23%
BioSR, training vs. test sources7.86% vs. 9.08%
BioSR, earlier vs. frozen calibration of PSF, gain, and noise6.4% vs. 7.9%
Disjoint halves of the training setat most 0.13 points

We appreciate your time and valuable feedback. We hope our responses address your concerns, and we would be grateful if you could reconsider your score.

Dear Reviewer hgSa,

Thank you again for your thoughtful review. We’ve posted a reply to your questions about the practical use of prior share and its sensitivity to the choice of source data. When you have a chance, we’d be grateful if you could take a look. If you have any further questions, we’d be happy to discuss them. If our reply addresses your concerns, we’d appreciate it if you could reconsider your rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Reply to the Area Chair and Reviewers

We thank the Area Chair for considering our submission and the reviewers for their time and detailed feedback. Several questions came up across the reviews, especially about how configurations were selected, the weak-lensing test set, the evidence on real observations, and the basis invariance of the prior share. We have marked the revisions in blue in the manuscript and collected image galleries on an anonymous page, available here, so reviewers can easily view the results while preserving double-blind review. Below, we summarize these shared concerns and how we addressed them in the revision.

  • Configuration selection (Reviewers 7uTy, wbgo). We re-ran selection on multiple validation images for weak lensing, Galaxy, and BioSR. The resulting rankings are consistent with the conclusions in the paper (Table 3, Appendices I.1 to I.3).
  • Weak-lensing test set. The test set excludes near-flat patches. On patches drawn without this filter, the ranking by physical error matches Table 3 (Appendix I.1).
  • Real observations (Reviewers hgSa, wbgo). On real BioSR observations, the solver with the smallest residual had the lowest PSNR and twice the zero-fill error in the unobservable set, consistent with the findings on simulated data. The benchmark measures this error directly to guide the choice among existing solvers, while the prior share quantifies the same risk before any data are acquired to inform measurement design (Fig. 6, Table 6, Appendices I.6 and I.9).
  • Error beyond the mean level. With an added operator that measures the mean level, one solver still had the third-lowest residual yet the second-lowest PSNR, with 95% of its error in the remaining unmeasured frequencies, confirming the finding in Section 5.2 that a small residual can hide a large error in unmeasured directions (Appendix I.4).
  • Basis invariance of the prior share (Reviewer wbgo). We clarified in Section 2 that within repeated-singular-value subspaces, aligning the singular directions with the eigenvectors of the source second moment makes the prior share strictly basis-invariant, matching our reported Galaxy and BioSR shares within 0.2% (Section 2, Appendix I.5).

We address the remaining questions and provide additional details in our individual replies to each reviewer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.