acceptodds
Withdrawn from ICLR 2027

When Minimizing the Measurement Residual Selects the Wrong Guidance Strength

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

When a diffusion or flow prior solves an inverse problem, its trajectory is pulled toward the measurement at every step, and the strength of that pull is a free parameter. Reported scores tune it against reference images; without one, the natural choice is the strength that fits the measurement best. We measure when that choice fails. Under released CIFAR-10 and face diffusion priors, the measurement residual falls as the strength grows, all the way to the strongest setting on every CIFAR-10 blur and mask. Behind binary masks, that setting is also the most accurate. Where the operator measures some directions only weakly, near its noise floor, selecting on the residual fits noise in those directions and gives up 0.099 to 0.514 in correlation and 0.039 to 0.118 under a latent prior. The fraction of such directions, the weak band, is read from the operator and the noise level before any prior is fitted. It tracks the cost. On 10 accelerated-MRI and compressed-sensing forward operators added afterward, it separated the costly operators, which gave up 0.184 to 0.475, from the rest, which gave up at most 0.016. The discrepancy principle, which stops when the residual reaches the noise level, gives up at most 0.028 where the band is not empty. Projected SURE, which estimates the error in all but the weakest directions, gives up at most 0.001 across all 28 operators. Both require the noise level. Most of the loss belongs to the correction rule that replaces the current estimate; the gradient rule stops where its step begins to overshoot. In the real-data task of predicting the brain's response to stimulation with a conditional flow prior, the residual falls the same way, but neither criterion's cost can be distinguished from zero.

Primary area:
generative models
Keywords:
diffusion models, inverse problems, posterior sampling, regularization parameter selection, discrepancy principle

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_C3HyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks how to choose the guidance strength ρ in diffusion-based inverse problem solvers when there is no reference image. The obvious choice is the ρ that fits the measurement best (lowest residual). The paper shows that this choice can go badly wrong. It defines a "weak band": the share of directions that the operator measures only weakly compared to the noise. The band can be computed from the operator and the noise level alone, before running any model. The paper claims it predicts when picking by residual is costly. It also tests alternatives that need no reference image: the discrepancy principle, SURE, GCV, and projected SURE. The evaluation covers blur and masks on cifar10, super-resolution on FFHQ faces (pixel and latent prior), MRI and compressed sensing (28 operators in total), and a real brain stimulation dataset.

Critical strengths

S1: The experiments are done carefully. Settings are chosen on half of the images and scored on the other half; the band thresholds were fixed in advance, and MRI and compressed sensing were added afterward as a test. The limitations are well stated.

S2: The practical advice is simple and useful. If the noise level is known, the discrepancy principle or projected SURE gives results close to the best setting.

Critical weaknesses

W1: The paper cites no diffusion solver that selects ρ by the lowest residual. Is this a real practice? If no method does this, it is not clear who the result is a warning for.

W2: The main result looks like a known one. In classical regularization, it is well known that fitting the measurement too closely amplifies noise in weakly measured directions. It is also known that the discrepancy principle and projected SURE avoid this. The paper says itself that the trade-off is classical, and the weak band is close to the classical comparison of singular values with the noise level. So the contribution is mainly an empirical check that this carries over to diffusion priors. It is not clear to me that this is enough for ICLR. The check is also limited in scope. Sec. 7 shows that with the DPS gradient rule, the loss is indistinguishable from zero on 2 of 5 blurs and 0.028 to 0.074 on the others. The paper explains this by saying the gradient rule cannot reach a low enough residual, not that it makes a better trade-off. But this is only tested on the 5 blurs, and ΠGDM is not tested at all. So for now the strong result holds for the replacement rule only.

W3: The brain stimulation experiment does not add much to the main claim. Accuracy is below 0.1, both costs are indistinguishable from zero, and the band is not tested there. Yet it is listed among the main contributions (Sec 1) and given a section of its own.

MW1: The paper is hard to read. There are many numbers without context and unusual wording ("gives up", "floor", "target", "pays"). There is a missing number in Sec. 6 ("give up as much as , the worst...") and a missing citation "(?)" in Related Work.

MW2: In Sec. 8, the first half (the released prior beats shrinkage when ρ is tuned) is a useful check. The second half, a flow prior trained by the authors that loses to shrinkage in 35/35 settings, does not test any selection criterion, and the paper itself calls the comparison uncontrolled. It is not clear what it adds to the main claim.

MW3: Each operator is scored on only 32 images, and most operators use cifar10 images, which are very small (32×32).

Questions for authors

Basically, to be accepted, the main weaknesses should be fixed. I’ll rephrase what I already pointed out in the weaknesses: (W1): Show which methods choose ρ based on the lowest residual. (W2): Explain what is new compared to the classical result. Test ΠGDM, and test DPS on more than the 5 blurs, or limit the claim to the replacement rule. (W3): Move the stimulation experiment to the appendix, or show that it supports the main claim.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer C3Hy, part 1 of 3

Thank you for the review. We answer each weakness in order, followed by the three requests under Questions for Authors. Section, figure, table, and appendix numbers refer to the revised PDF. Image costs are losses in correlation on the scoring half, relative to the oracle strength chosen by accuracy on the selecting half.

W1. Who selects the strength by the lowest residual. ZAPS (Alçalar & Akçakaya, ECCV 2024) tunes the per-step weights of DPS by minimizing ∥y−Ax^0∥2\|y - A\hat{x}_0\|^2 over all measurements. Solvers whose data steps enforce exact consistency sit at the residual's minimum by construction, as in the range-null-space step of DDNM without noise, the replacement of known pixels in RePaint, and the projection step of score-based MRI. Other methods stop at a noise-level threshold (ReSample, SITCOM), adapt the data weight by SURE (SMRD), or predict held-out measurements (ZADS). Sections 1 and 3 cite these methods, and Sec. 3 states that our residual criterion, discrepancy principle, and SURE apply the rules of ZAPS, of ReSample and SITCOM, and of SMRD, respectively, on a common sampler with one strength per trajectory. ZADS, by the authors of ZAPS, gives its reason for holding out data (its Sec. II-D): "While such a strategy is effective for DPS-type sampling, as it applies mild corrections that remain close to the noisy manifold, the stronger CG-based updates in DDS may lead to overfitting if the same measurements are used for both inference and supervision, necessitating the use of hold-out masking in the current setup." Our results show the operators on which this overfitting costs accuracy, and Sec. 8 compares the three data steps. The result is a warning for users of residual-based tuning, such as ZAPS, and for users of exact data-consistency steps, and the paper also compares the alternatives available to them. Appendix U discusses recent residual-constrained solvers (CDIM, FAST-DIPS), sampler parameter design by Benita et al., GCV and SURE tuning of plug-and-play methods, and two concurrent preprints.

W2 (a). What is new beyond the classical result. The trade-off is classical, as the paper states. Table 2 and Appendix I apply the same five criteria to Tikhonov shrinkage on the same 28 operators, images, and halves. Two conclusions transfer to the sampler with the pixel priors and the replacement rule. Where the band is not empty, the residual criterion costs 0.060 to 0.936 under shrinkage and 0.099 to 0.514 under the sampler. There, the discrepancy principle costs at most 0.034 and 0.028, respectively, and projected SURE with one probe costs at most 0.002 and 0.001. Two other results that hold for shrinkage do not transfer. SURE of the measurement error costs at most 0.026 under shrinkage, even with the one-probe divergence used for the sampler. Under the sampler, it picks the residual criterion's strength on 3 of the 8 CIFAR-10 blurs and graded masks, at a cost of up to 0.514. The true measurement error picks SURE's strength on 7 of the 8 and also costs up to 0.514 in correlation and 16.2 dB in PSNR. Across the 32 selecting images, the mean one-probe SURE estimate lies within 7.3 percent of the mean true error at every strength (2.9 percent with four probes). Minimizing even the true measurement error can select poorly (Sec. 7, App. O). Under the latent prior, the residual stays above the noise level, so the discrepancy principle repeats the residual criterion's choice on all six operators, at a cost of up to 0.118.

We also agree that the band rests on the classical comparison of singular values with the noise level, and Sec. 2 cites Hansen (2010) for it. We turn this comparison into a check that runs before a nonlinear sampler is used, with thresholds that we fixed in advance. Its correlation with the cost is 0.890 on the 18 operators measured first and 0.805 on all 28, including the 10 MRI and compressed-sensing operators added afterward (Sec. 6). On LSUN churches and bedrooms, it is 0.911 and 0.889 (App. S). Weighting the directions by image content instead, as a signal-to-noise ratio does, breaks the relation (correlation -0.222, App. F). Section 10 now opens with this summary of what carries over.

AuthorsOpenReview ↗

Reply to Reviewer C3Hy, part 2 of 3

W2 (b). Testing ΠGDM and testing DPS beyond the five blurs. We ran ΠGDM and DPS on all 18 operators on CIFAR-10 images and on three 256 × 256 face super-resolution operators (Sec. 8, App. N, Table 5). We ran ΠGDM as in Algorithm 1 of its paper, for the variance-preserving case with η = 0, and swept its strength from 0.03 to 100, where strength 1 gives the published update. Both rules use the images, halves, samples, steps, and seeds of the replacement sweeps and run on our deterministic sampler (100 steps on CIFAR-10 and 50 on faces) instead of the 1000 ancestral steps of DPS (Secs. 2 and 10).

Under ΠGDM, minimizing the residual costs 0.006 to 0.426 on the 11 operators with a band on CIFAR-10 images, with every interval excluding zero, against 0.226 to 0.514 under the replacement rule. The failure also occurs without a proximal step. On the three face operators, minimizing the residual under ΠGDM costs 0.206 and 0.246 on the two with a band and 0.061 on the box prefilter, whose band is empty. Under DPS, it costs 0 to 0.295 on the CIFAR-10 operators with a band, with intervals excluding zero on 6 of the 11, and at most 0.006 on the face operators. On the tested grids, DPS reaches a higher minimum residual than the replacement rule on all 18 operators on CIFAR-10 images. Where their residuals overlap, the two rules agree in accuracy to within 0.085. DPS's smaller cost accompanies a higher minimum residual with these 100 steps. We did not vary the step schedule.

Under ΠGDM, the residual turns back up past its minimum. On three operators, the residual closest to the noise level lies past the minimum, so the discrepancy principle selects a strength there and costs up to 0.377. Restricted to strengths up to the minimum, it costs at most 0.060 under ΠGDM and 0.028 under DPS. We chose this restriction after seeing these failures. We ran the three face operators after fixing the restricted variant, and on all three it picked the most accurate strength of ΠGDM.

W2 (c). Risk estimates under the gradient rules. We also ran SURE, GCV, and projected SURE under both rules on the 18 operators on CIFAR-10 images, with the probes of the replacement sweeps (Sec. 8, App. N, Table 6). Projected SURE costs up to 0.377 under ΠGDM and 0.329 under DPS, with intervals excluding zero on 6 and 8, respectively, of the 11 operators with a band. In all 14 of those cases, it selects at or past the residual's minimum. A probe-size check shows that, under these rules, the change a probe makes does not grow in proportion to its size, and the measurement-space divergence estimate depends on the size. Under the replacement rule, multiplying the probe size by 100 multiplies the median absolute pixel change by 94 to 108, and the divergence estimates agree to within 3.2 percent. Under the gradient rules, the median growth factor across cases is 1.45 at the residual's minimum and 1.04 at the next stronger strength. At that stronger strength, the measurement-space divergence estimate changes sign across the three probe sizes in 41 of 72 cases. The estimates varied over the tested range, and we did not establish that they converge at smaller sizes. These tests do not establish why projected SURE selects poorly there (App. N). Under the latent prior, the estimates also depend on the probe size, so we did not run the risk estimates there (App. H).

W2 (d). Scope of the safety claim. The abstract and the main finding in Sec. 1 limit the spectral safety claim to the replacement rule. Step 2 of Sec. 9 requires both an empty band and smin⁡≥3.21s_{\min} \ge 3.21, a boundary observed on the tested operators. It also reports the counterexample under ΠGDM (0.061 on the face box prefilter, whose band is empty and whose smin⁡s_{\min} is 10). The failure on operators with a band extends to ΠGDM and partly to DPS, as reported above.

AuthorsOpenReview ↗

Reply to Reviewer C3Hy, part 3 of 3

W3. The stimulation experiment. We moved the stimulation experiment to Appendix Q and removed it from the contributions. Section 10 mentions it in one sentence. Some released electrode tables list a contact twice, and our field model counted that contact twice in 10 of the 91 runs. After we corrected this double count, both held-out costs remain indistinguishable from zero: 0.028 [-0.021, 0.067] for the residual criterion and 0.009 [-0.031, 0.060] for the discrepancy principle. The conclusion about the decline past the pooled peak (ρ = 30) changed: accuracy now falls by 0.037 [0.009, 0.061], with an interval excluding zero. The band is not shown there because it is read in the coordinates of a fitted prior and cannot be computed before fitting that prior.

MW1. Readability, wording, and the missing number and citation. We rewrote the paper for clarity: sentences are shorter, each term is explained where it first appears, and each results section opens with a boxed finding stated in words. The words gives up, floor, target, and pays no longer appear, and floor margin is now smin⁡s_{\min}, defined in Sec. 2 as the smallest sis_i above 0.1. The missing number and the "(?)" came from compiling the submitted PDF with outdated macro and bibliography files. The number was 0.514, and the citation was Xia et al. (2026), NA-NHMC, which Sec. 3 now cites. The revised PDF has no missing numbers or citations.

MW2. The fitted flow. We moved the fitted-flow comparison to Appendix L, and no claim about selection rests on it. We retained the released-prior comparison as a check: after both methods are tuned on validation images, the released CIFAR-10 prior beats Tikhonov shrinkage on all five blurs (Sec. 4, App. L).

MW3. The number of scored images and the small CIFAR-10 images. On four operators, we now use 256 images and score 128 of them. Every selected strength is unchanged, and each cost moves by at most 0.013 (App. P). Ten of the 28 operators (six face super-resolution and four MRI operators) use 256 × 256 FFHQ images, as do the latent-prior experiments (Table 1). We also repeated the six face super-resolution operators on LSUN churches and bedrooms at 256 × 256, with their released priors and 64 validation images each (Sec. 5, App. S). There, minimizing the residual costs 0.081 to 0.247 on the operators with a band, with every interval excluding zero, and at most 0.043 on those without. The discrepancy principle costs at most 0.008, and projected SURE chooses the oracle's strength in all 12 cases. In each of 1000 random splits of the images, every binary mask costs exactly zero, and every pixel-prior operator with a band costs a positive amount (Sec. 5, App. A). As Sec. 10 states, all operators are still scored on 32 images, and four of them are also scored on 128.

Questions for Authors. For W1, Sec. 1 names ZAPS, which tunes the strength by minimizing the residual, and the solvers whose exact data-consistency steps sit at the residual's minimum. For W2, Table 2 shows what is new compared with the classical result, and Secs. 7 and 10 explain it. We also tested ΠGDM and DPS on all 18 operators on CIFAR-10 images and on three face operators, and we limited the safety claim to the replacement rule. For W3, the stimulation experiment is now in Appendix Q.

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

This paper studies how to choose guidance strength when using diffusion or flow priors to solve inverse problems without access to the original signal. The authors explore when selecting the strength that best fits the measurement leads to worse reconstruction. They analyse different indicators, operators, and noise levels, and investigate this failure across different priors and correction rules. They further compare existing selection criteria, including the discrepancy principle and projected SURE, and examine a brain stimulation prediction task.

Critical strengths
  1. The paper tackles an important and useful problem: choosing guidance strength when the original signal is unavailable.

  2. The analysis covers several measurement operators, generative priors, and application settings, providing useful comparisons of when the failure occurs.

Critical weaknesses
  1. Clarity and accessibility.

The paper is extremely difficult to follow without substantial background in this specific field. The problem, basic terminology, and assumptions need clearer explanations before the technical analysis. The presentation is heavy with numerical results, notation, and experiments, but provides insufficient intuition. Much of the discussion describes results without clearly explaining their implications or giving a clear takeaway from each analysis.

The abstract similarly emphasizes technical details and numerical ranges without clearly communicating the motivation and contribution. A method overview figure, introductory preliminaries, and an earlier explanation of the relationship to prior work would substantially improve accessibility (Questions 1 and 2).


  1. Limited methodological contribution and validation of the practical conclusions.

The empirical analysis is useful, but the additional contribution beyond established regularization principles and selection criteria remains insufficiently clear. The paper does not establish sufficiently general guidance about which criterion should be used in which circumstances. The conclusions depend on the correction rule, experimental setup, and noise assumptions, limiting their practical applicability. The requirement for a known noise level is also restrictive, while the small number of evaluation images limits confidence in the generality of the findings. Stronger validation and clearer practical conclusions are needed to establish the significance of the contribution (Questions 3 to 6).

Questions for authors
  1. Problem definition and overview. Can the authors introduce the problem, basic terminology, and selection criteria more intuitively before the technical analysis? An overview figure and clearer preliminaries would help readers understand the setup.

  2. Interpretation and takeaways. Can each main analysis explain why the reported pattern occurs, what it establishes, and what practical conclusion follows? The current presentation often lists numerical results without making these implications clear.

  3. Additional contribution. What new understanding does the paper establish beyond classical methods? Which findings constitute the main contribution?

  4. Generality of the conclusions. Which conclusions hold across correction rules, priors, operators, and noise settings, and which are specific to the tested configurations? Can the authors provide clearer guidance about when each selection criterion should be used?

  5. Noise level requirements. How practically useful are the recommended criteria when the noise level must be estimated? Beyond the reported sensitivity checks, what evidence supports their reliability under realistic noise estimation errors or departures from the assumed noise model?

  6. Evaluation size. The main image experiments use only 64 images per domain, with 32 used for scoring. Do the conclusions hold on substantially larger and more diverse evaluation sets?

4 comments
AuthorsOpenReview ↗

Reply to Reviewer wTjL, part 1 of 3

Thank you for the review. We answer the two weaknesses and the six questions in order. Section, figure, table, and appendix numbers refer to the revised PDF. Image costs are losses in correlation on the scoring half, relative to the oracle strength chosen by accuracy on the selecting half.

W1. Clarity and accessibility. We rewrote the paper for readers outside this subfield. Section 2 now defines the measurement model, guidance strength, residual, weak band, and selection criteria before the first result. Figure 1 gives an overview, Algorithm 1 writes out the sampler, and related work appears earlier, in Sec. 3. Each results section (Secs. 5 to 8) opens with a boxed finding, followed by the evidence and an explanation. The abstract starts with the motivation: published results tune the strength with ground truth, which a deployed solver lacks. It then states the question, the results, and the tested conditions in words and keeps only three numbers (details under Q1 and Q2).

W2. Contribution and practical conclusions. Q3 compares the findings with classical regularization, Q4 states when each criterion worked, Q5 covers noise estimation and other noise models, and Q6 reports the larger image sets and the scene experiments. For these questions, we added a classical baseline on the same problems, ran the gradient rules of ΠGDM and DPS on 21 operators, tested three other noise models, scored more images, and repeated the face super-resolution operators on two more image domains. Each conclusion is now stated together with the rules, priors, and noise models on which it was tested.

Q1. Problem definition and overview. Section 2 introduces the problem before the first result. We observe y=Ax+ηy = Ax + \eta with a known operator AA and white Gaussian noise η\eta of level σ. A guided sampler alternates a denoising step with a data step whose strength sets how closely the estimate is fitted to yy. The residual ∥y−Ax^∥\|y - A\hat{x}\| measures that fit and can be computed without ground truth. The weak band is the fraction of image directions along which a unit change of the image moves the measurement by more than 0.1 and at most 1 noise standard deviation. It is computed from AA and σ alone. The residual criterion picks the strength with the lowest residual, and the discrepancy principle picks the strength whose residual is closest to the noise level. SURE estimates the measurement error ∥A(x^−x)∥2\|A(\hat{x} - x)\|^2, generalized cross-validation divides the mean squared residual by a factor computed from the divergence of the fit, and projected SURE estimates the image error in directions with si>0.1s_i > 0.1. Each of these three is minimized over the strength grid. Figure 1 shows one sampler step, the spectra of three operators with their bands, and what each criterion selects on one blur. Table 1 lists the six operator families.

Q2. Interpretation and takeaways. Each results section opens with its finding and then gives the reason for it.

  1. Section 5 explains the noise amplification: as ρ grows, the data step approaches a least-squares fit. In each weakly measured direction, this adds coefficient noise with a standard deviation from 1 to less than 10. The measured errors agree with this explanation (App. D). On the CIFAR-10 blurs and graded masks, the band holds 0.492 to 0.836 of the squared error added by minimizing the residual.
  2. Section 6 shows that the band predicts the cost, because the cost follows the noise that a least-squares fit injects into the weakly measured directions. Weighting each direction by how much of the image it carries, as a signal-to-noise ratio does, breaks the relation (App. F).
  3. Section 7 explains why SURE of the measurement error can select poorly. The added error lies mostly in the weak band, and the measurement error weights squared image errors by ςi2=σ2si2\varsigma_i^2 = \sigma^2 s_i^2, which gives those directions little weight. The divergence term of projected SURE instead includes weights of 1/ςi21/\varsigma_i^2. Generalized cross-validation with one probe fails for another reason: as the fit becomes exact, both its residual and denominator approach zero, so probe error can dominate the denominator.
  4. Section 8 shows that increasing strength past the lowest-residual setting can raise the residual again under the gradient rules. The discrepancy principle can select a strength past that minimum, and projected SURE can select poorly there too.

Section 9 turns the findings into a four-step procedure, and Appendix T (Table 11) lists every diagnostic study with what it shows.

AuthorsOpenReview ↗

Reply to Reviewer wTjL, part 2 of 3

Q3. Additional contribution. To separate what is classical from what the sampler changes, Table 2 and Appendix I apply the same five criteria to Tikhonov shrinkage on the same 28 operators, with the same images and the same split into halves. Two classical conclusions carry over to the sampler with the replacement rule and the pixel priors. Minimizing the residual fails where the weak band is not empty (costs of 0.060 to 0.936 for shrinkage and 0.099 to 0.514 for the sampler), and the discrepancy principle and projected SURE avoid the failure (worst costs of 0.028 and 0.001 under the sampler). The paper adds four findings beyond these.

  1. SURE of the measurement error costs at most 0.026 for shrinkage but up to 0.514 under the sampler, where it picks the residual criterion's strength on 3 of the 8 CIFAR-10 blurs and graded masks. The true measurement error picks SURE's strength on 7 of the 8 and also costs up to 0.514. Across the 32 selecting images, the mean one-probe SURE estimate lies within 7.3 percent of the mean true error at every strength. Minimizing even the true measurement error can select poorly (Sec. 7, App. O).
  2. Under the latent prior, the residual never falls to the noise level, and the discrepancy principle repeats the residual criterion's choice on all six operators, at a cost of up to 0.118 (Sec. 7, App. H).
  3. Under the gradient step of ΠGDM, minimizing the residual costs accuracy on every tested operator with a band (up to 0.426), and under DPS it costs less (up to 0.295). On the 18 operators on CIFAR-10 images, projected SURE also fails under both rules (up to 0.377). In all 14 cases where its cost interval on an operator with a band excludes zero, it selects at or past the residual's minimum (Sec. 8, App. N).
  4. The weak band works as a check before sampling. Its thresholds were fixed before we added 10 MRI and compressed-sensing operators, and on those operators it still separated the costly operators from the rest (Sec. 6).

Q4. Generality of the conclusions and when to use each criterion. Section 10 states the scope of each conclusion, and Sec. 9 gives the procedure.

  1. With the replacement rule and white Gaussian noise, minimizing the residual fails where the band is not empty, and the discrepancy principle and projected SURE avoid the failure. This held for both pixel priors, at both resolutions, in all six families, and for super-resolution on LSUN churches and bedrooms with their own released priors (App. S).
  2. Under the latent prior, the failure carries over, and the discrepancy principle repeats it (App. H).
  3. Under ΠGDM, minimizing the residual failed on all 13 tested operators with a band and also cost 0.061 on one face operator without a band. Section 9's safety rule requires an empty band and smin⁡≥3.21s_{\min} \ge 3.21 and applies to the tested operators under the replacement rule. Under DPS, minimizing the residual cost less (up to 0.295).
  4. Under heavy-tailed, signal-dependent, and correlated noise on four CIFAR-10 operators, the residual criterion picks the same strengths as under white noise in all 12 cases, and the discrepancy principle costs at most 0.024. Projected SURE costs up to 0.082 under correlated noise (App. R).

Section 9 then gives four steps.

  1. Before sampling, compute the band and smin⁡s_{\min} from AA and σ.
  2. Under the replacement rule, if the band is empty and smin⁡≥3.21s_{\min} \ge 3.21, minimizing the residual is safe, and on our operators it cost at most 0.017. The boundary 3.21 was observed on the tested operators and is not a guarantee for others.
  3. Otherwise, do not minimize the residual. With a pixel prior and a trusted noise level, use the discrepancy principle. Under the replacement rule, projected SURE is an alternative if a second reconstruction per setting and the pseudo-inverse are affordable and the noise is white. If the residual turns back up as the strength grows, apply the discrepancy principle only up to the residual's minimum. We chose this variant after seeing the CIFAR-10 sweeps and then tested it on three face operators, and on all three it picked the most accurate strength of ΠGDM.
  4. If the noise level is uncertain, calibrate it or read it from the observation (Q5).
AuthorsOpenReview ↗

Reply to Reviewer wTjL, part 3 of 3

Q5. Noise-level requirements. The residual criterion and generalized cross-validation need no noise level. For the criteria that do, we report four analyses.

  1. A misjudged level (Sec. 7, Figure 4b). Equation 1 depends on ρ and σ only through ρ/σ2\rho/\sigma^2, so a misjudged level is equivalent to a rescaled strength grid. Within 10 percent of the true level, the worst cost over the first 18 operators is 0.294 for the discrepancy principle (on the one operator whose residual barely reaches the noise level) and 0.074 for projected SURE. An underestimate does more harm: at 0.75 times the true level, the two criteria cost up to 0.353 and 0.514, and at half the true level, both reach the residual criterion's cost of 0.514. Overestimates of up to a factor of two cost them at most 0.322 and 0.190.
  2. The level read from the observation (App. J). On 9 of the 28 operators (four blurs, one Gaussian prefilter, and all four MRI operators), we estimate σ from measurement directions with little or no signal, without using the true level. The estimates lie between 0.93 and 1.08 times the true level (within 1.1 percent for MRI). With them, the discrepancy principle selects the same strength as with the true level on all 9, at a cost of at most 0.016. This tests selection among reconstructions computed with the true level. At strength ρ, a data step using σ^\hat{\sigma} would return the true-level reconstruction at ρ(σ/σ^)2\rho(\sigma/\hat{\sigma})^2, with the factor between 0.85 and 1.16 here, while adjacent strengths on our grid differ by a factor of about 3. We did not test that complete procedure with the estimated level. Masks, box prefilters, and Gaussian matrices leave no measurement direction without signal, so their level has to come from calibration, as with the noise-only scan in MRI.
  3. The band tolerates errors in σ on these operators (App. E). Since si=ςi/σs_i = \varsigma_i/\sigma, a level that is off by a factor of cc moves both thresholds by the same factor. With σ off by a factor of 0.5, 2, or 3, the band's correlations with the cost of the residual criterion are 0.789, 0.893, and 0.861, respectively, against 0.890 at the true level.
  4. Other noise models (App. R). On four CIFAR-10 operators (the blurs of width 0.8 and 1.2, the half mask, and the graded mask that weights down 60 percent of the kept pixels), we replaced the Gaussian white noise with Laplace noise, signal-dependent noise (a Gaussian approximation to Poisson-Gaussian sensor noise), and spatially correlated noise (adjacent entries correlated at 0.5). Each has the same mean variance, and every criterion receives the nominal σ. The residual criterion selects the same strength as under Gaussian white noise in all 12 cases, and the discrepancy principle costs at most 0.024 (0.021 under Gaussian white noise). Projected SURE costs at most 0.001 under the first two models but 0.082 on the graded mask under correlated noise, which its derivation does not cover. Selecting well under the first two models does not establish that the Gaussian risk formula is unbiased there. SURE and generalized cross-validation select poorly under every model.

Q6. Evaluation size. We added two experiments and one check.

  1. More images (App. P, Table 8). For a blur, a binary mask, a graded mask, and compressed sensing, we reconstructed 192 more CIFAR-10 test images, selected the strength on 128 of the 256 images, and scored it on the other 128. All four operators keep the strengths selected on 64 images, each cost moves by at most 0.013, and the intervals of the residual criterion's cost narrow 2.6- to 2.8-fold.
  2. Other image domains (App. S, Table 10). We repeated the six face super-resolution operators on LSUN churches and bedrooms at 256 × 256, with the released priors for those scenes and 64 validation images each. Where the band is not empty, minimizing the residual costs 0.081 to 0.188 on churches and 0.126 to 0.247 on bedrooms, with every interval excluding zero, against 0.099 to 0.234 on faces. Where the band is empty, it costs at most 0.043. With its thresholds fixed before these runs, the band's correlation with the cost is 0.911 on churches and 0.889 on bedrooms. The discrepancy principle costs at most 0.008, and projected SURE selects the oracle's strength in all 12 cases.
  3. Other splits (Sec. 5, App. A). In each of 1000 random splits of the 64 images, every binary mask costs exactly zero, and every pixel-prior operator with a band costs a positive amount. The central 95 percent range of the cost of each operator with a band is no wider than 0.063.

All operators are still scored on 32 images, and four of them are also scored on 128, as Sec. 10 states.

I thank the authors for their detailed responses and the substantial additional analyses. The expanded experiments and clearer statements about the scope of the conclusions substantially address Q4 and Q6. The noise analyses also provide a meaningful response to Q5, although validation of the complete procedure using estimated noise remains limited.

However, my main concern about clarity remains unresolved. The revised manuscript and rebuttal are still very heavy with technical details, numerical results, and references to different experiments. While definitions, figures, and explicit findings have been added, these additions do not sufficiently explain the intuition or make the central argument easy to follow. In several responses, the details obscure the answer rather than clarify what the findings establish and why they matter. For example, the new overview figure does not help clarify the approach. It is visually difficult to follow, contains visual artifacts, and, in my opinion, adds confusion rather than improving understanding. The introduction also remains difficult to follow, as do the dense formulations and explanations. The bullet points summarize the main findings in a sentence, but in many cases this is insufficient to make the reasoning and implications clear.

Moreover, regarding Q3, the additional comparisons better identify the claimed empirical contribution, particularly the differences between classical and generative reconstruction. Nevertheless, I remain unconvinced that the findings establish sufficiently important new understanding and practical impact beyond the established principles discussed in the paper. The additional evidence strengthens the analysis, but does not fully resolve this concern.

Overall, the authors have provided meaningful additional analyses and explanations, which I appreciate. However, the clarity issue remains severe in my opinion. Together with the remaining concerns, this leads me to tentatively maintain my weak rejection recommendation. I will nevertheless follow the discussion with the other reviewers and consider their assessments and remaining concerns before making my final recommendation.

Reviewer_L6KhOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

There is no actual claim, not a provided method, no satisfactory empirical results, no novelty, and no contribution to this paper. The text is filled with misleading and problematic statements. Some included in the weaknesses section, however, this extends to all of the paper. The setting is not clear. The inverse problems are not clearly explained. There is only one visual qualitative image, despite claims of other inverse problems such as accelerated MRI.

Critical strengths

None exists.

Critical weaknesses

Language of the paper seems misleading and inaccurate in many cases.

  1. "Reported scores tune it against reference images; without one, the natural choice is the strength that fits the measurement best." This is problematic in many levels. Depending on how the generative image solver is designed (MMSE, MAP, Posterior sampling), the objective consists of two terms, data-fidelity and regularizer. Generative inverse solvers don't just minimize the data fidelity, they minimize the whole objective function, including data-fidelity + regularizer. Thus, the natural choice mostly is not most faithful solution to the measurement.

  2. "We measure when that choice fails. Under released CIFAR-10 and face diffusion priors, the measurement residual falls as the strength grows, all the way to the strongest setting on every CIFAR-10 blur and mask." The contribution is based on the hypothesis that is wrong! First, the best solution is not again the most faithful one. So measuring when this does not happen is literally every case of generative inverse solver. Second, it is not clear what the authors means by blur and mask, what kind of inverse problem is being investigated here?

  3. "Behind binary masks, that setting is also the most accurate. Where the operator measures some directions only weakly, near its noise floor, selecting on the residual fits noise in those directions and gives up 0.099 to 0.514 in correlation and 0.039 to 0.118 under a latent prior." This is extremely unclear. What setting? What binary mask? What is the inverse problem? What do the author means by operator at the noise floor? What noise floor? Correlation with what?

  4. Same goes for the rest of the abstract. The claim and contribution of the paper is not at all clear. The inverse problems and method is muddled.

  5. Introduction first paragraph is not cited at all, with very vague and inaccurate claims. One example is "the clean signal is pulled toward the measurement"; unfortunately, in the case of inverse problems, there is no clean signal. Pulling towards the measurement is not the correct term used. Either projection toward data-consistent manifold or gradient descent of data-fidelity is used.

  6. Intro second paragraph: we don't dont have references in any inverse problem, literally!

  7. "When the measurement can be split, the strength follows from predicting a held-out part, as ZADS does for accelerated MRI (Alçalar et al., 2025). Holding out entries replaces A with a row subset whose spectrum differs from that of the full operator; we do not study that case." These statement are extremely inaccurate.

  8. Unfortunately, this paper has no actual contribution. Looking at contribution list, none counts as actual contribution one can present in the literature of generative inverse problems. 3 and 6 are actually definition or setting of inverse problems, there is no reference available.

  9. None existing references: (Song et al., 2024)!

  10. Sampling rule in Eq.1 is wrong. The paper claims ρ\rho is the guidance strength, which is simply not true. This update rule has not renoising common to diffusion model and it is not clear how the mean of denoiser (diffusion at a certain noise level) is update! That is simply output of diffusion and can not be updated manually!

There is simply no novelty, no contribution to this paper. There is no claim, not enough consistent empirical presented. The paper claims to be investigating accelerated MRI, yet no evidence is presented.

Questions for authors

NA

3 comments
AuthorsOpenReview ↗

Reply to Reviewer L6Kh, part 1 of 3

Thank you for the review. We answer each point below, using the review's numbering. Section, figure, table, and appendix numbers refer to the revised PDF. Image costs are losses in correlation on the scoring half, relative to the oracle strength chosen by accuracy on the selecting half.

Summary (claims, method, results, and contribution). The paper studies how to choose the guidance strength of diffusion solvers for linear inverse problems when no ground-truth image is available. Section 1 lists four findings and identifies the evidence for each. The main finding is that, with the data step of Eq. 1 (the replacement rule), the operator and the noise level indicate before sampling whether minimizing the measurement residual is safe on the operators tested. The main pixel-prior study covers 28 operators in six families and five criteria that need no reference image. We also test a latent prior on six super-resolution operators and compare three data steps on 21 operators. Every criterion selects a strength on half of the images and is scored on the other half (Sec. 2).

Summary (setting and inverse problems). Section 2 now defines the problem before the first result. We observe y=Ax+ηy = Ax + \eta, where the operator AA is known and η\eta is white Gaussian noise of level σ, and we reconstruct the image xx. Table 1 lists the six families (deblurring, inpainting with binary masks, inpainting with graded masks, super-resolution, accelerated MRI simulated on face images, and compressed sensing), and Appendix B defines all 28 operators.

Summary (one qualitative image). Figure 3 shows one image from the scoring half for each of six operators: a blur, a graded mask, a binary mask, face super-resolution, 8-coil MRI, and compressed sensing. Each image is reconstructed at the strengths chosen by the oracle, the residual criterion, the discrepancy principle, and projected SURE. Appendix M (Figures 6 to 8) repeats the figure for the next three images of the scoring half, taken in order. The MRI results are listed in our reply to your final paragraph.

Summary and opening of the weaknesses (misleading language). We rewrote the text in standard terms and added citations for the methods it describes. Points 1 to 10 list the corrections.

1. Data fidelity and the regularizer. At a fixed strength, the solver balances the data-fidelity term against the prior, as Sec. 2 now states. The strength sets this balance and must itself be chosen. The paper asks how to choose it without ground truth. The residual is then a directly observable measure of fit, and published methods minimize it. ZAPS tunes the per-step weights of DPS by minimizing ∥y−Ax^0∥2\|y - A\hat{x}_0\|^2 over all measurements (Alçalar & Akçakaya, 2024). Other solvers sit at the residual's minimum by construction because their data steps enforce exact consistency, as in DDNM without noise, RePaint, and score-based MRI (Sec. 1). Calling the best-fitting strength "the natural choice" was too broad, and we removed that wording.

2. "Literally every case" and the meanings of blur and mask. The most faithful setting is often not the most accurate, and the paper measures the operators on which this happens. For all four binary masks, the strength with the lowest residual is also the most accurate, with a cost of exactly zero in each of 1000 random splits of the images. Among the main 28 pixel-prior operators, those without a weak band cost at most 0.033. Those with a band cost 0.099 to 0.514 (Finding 1, Sec. 5). By "blur" we mean a circular Gaussian blur with a width of 0.8 to 2 pixels, and the inverse problem is deblurring. By "mask" we mean either a random binary mask that keeps 15 to 70 percent of the pixels or one of three graded masks, which keep half of the pixels and give 30, 60, or 90 percent of the kept ones a weight between 0.15σ and 0.95σ. The inverse problem is then inpainting (Table 1, App. B).

AuthorsOpenReview ↗

Reply to Reviewer L6Kh, part 2 of 3

3. The abstract sentence on binary masks and the noise floor. A setting is a value of ρ on a fixed grid that runs from 10−410^{-4} to 1 for blurs and from 10−410^{-4} to 30 for the other operators, with two values per decade (Sec. 4). "That setting" meant the strongest value on the grid, which had the lowest residual on every CIFAR-10 blur and mask. The binary masks keep 15, 30, 50, or 70 percent of the pixels at full weight. The inverse problem is the recovery of xx from y=Ax+ηy = Ax + \eta for the 28 operators of Table 1. Section 2 replaces "noise floor" with a definition. With A=∑iςiuivi⊤A = \sum_i \varsigma_i u_i v_i^\top and si=ςi/σs_i = \varsigma_i/\sigma, a unit change of xx along viv_i moves the measurement by sis_i noise standard deviations. The weak band is the fraction of directions with 0.1<si≤10.1 < s_i \le 1. In these directions, a least-squares fit gives the coefficient of xx along viv_i an error with standard deviation 1/si1/s_i, from 1 to less than 10. Correlation is measured between the reconstruction and the true image, over all pixels and channels, on images not used for selection. Appendix G adds PSNR, SSIM, and LPIPS.

4. The rest of the abstract. The abstract first explains what the guidance strength does and why the residual is available without ground truth. It then asks when minimizing the residual selects the wrong strength and which classical criteria choose better without a reference image. It names the tasks and the noise model, defines the weak band in words, and gives the main results and their scope. The study uses the evaluation protocol of Sec. 2 and the check of Sec. 6, which runs before sampling. Section 9 gives a four-step procedure for choosing the strength.

5. Introduction, first paragraph. The paragraph now cites DDRM, score-based MRI, DPS, ΠGDM, DDNM, DiffPIR, and a survey, and it uses the two standard descriptions that you name. Gradient rules move the iterate down the gradient of the data-fidelity term. Proximal and projection rules replace the denoised estimate with the solution of a regularized least-squares problem centered on it or with its projection onto the images consistent with yy. The phrase "the clean signal" meant the denoiser's estimate x^0\hat{x}_0 of the clean image, and the text now says so.

6. Introduction, second paragraph. The paragraph now cites ZAPS, DDNM, RePaint, score-based MRI, ReSample, SITCOM, SMRD, DDS, and ZADS. The third paragraph cites classical work on choosing a regularization parameter, from Morozov (1966) and Golub et al. (1979) to Hansen (2010) and Lucka et al. (2018).

7. ZADS. We agree that the submitted sentences were inaccurate, and we removed them. Section 1 now says that ZADS holds out part of the k-space because the conjugate-gradient data step of DDS, on which it builds, corrects more strongly than DPS, and reusing the measurement for supervision risks overfitting (Alçalar et al., 2025). Section 3 states that we do not test held-out prediction, and Sec. 10 lists measurement splitting as future work.

AuthorsOpenReview ↗

Reply to Reviewer L6Kh, part 3 of 3

8. Contributions. The list now has four findings, each with a reference to its evidence, and the stimulation study is no longer among them. The definitions have moved to Sec. 2, with references. Comparing singular values with the noise level is standard in the analysis of discrete ill-posed problems (Hansen, 2010), and the paper cites the original source of each criterion. The study tests which classical conclusions carry over to guided samplers and finds two that do not. First, SURE of the measurement error, which is reliable for Tikhonov shrinkage on the same problems, picks the residual criterion's strength under the sampler on 3 of the 8 CIFAR-10 blurs and graded masks. Second, under the latent prior, the discrepancy principle repeats the residual criterion's choice (Sec. 7).

9. The reference (Song et al., 2024). This citation refers to ReSample, published at ICLR 2024 as "Solving inverse problems with latent diffusion models via hard data consistency" by B. Song, S. M. Kwon, Z. Zhang, X. Hu, Q. Qu, and L. Shen. ReSample is in the reference list, and Sec. 3 now names it.

10. The sampling rule in Eq. 1. After the data step, the sampler re-noises the corrected estimate to the next step's noise level with the DDIM update. The submission did not write this step out. Algorithm 1 now shows the full loop, with re-noising in line 6. The data step corrects the denoised estimate x^0\hat{x}_0, and the sampler forms the next iterate from it, as in DDNM, DiffPIR, and DDS. Equation 1 gives the minimizer of ∥y−Ax∥2/σ2+∥x−x^0∥2/vt\|y - Ax\|^2/\sigma^2 + \|x - \hat{x}_0\|^2/v_t with vt=ρ(1−αˉt)v_t = \rho(1 - \bar{\alpha}_t), so ρ sets how strongly the step enforces yy. As ρ → 0, the sample is unconditional, and as ρ → ∞, the step becomes the range-null-space step x^0+A+(y−Ax^0)\hat{x}_0 + A^+(y - A\hat{x}_0) of DDNM. At ρ = 1, vtv_t is the variance of ΠGDM's Gaussian approximation. At a given step, matching DiffPIR's data-step variance gives ρt=1/(λαˉt)\rho_t = 1/(\lambda\bar{\alpha}_t) (Sec. 2), a relation that changes along the trajectory. The term "guidance strength" describes the balance set by ρ, just as DPS's step size and DiffPIR's λ set that balance in their own updates. Appendix A checks the data step against the closed-form Gaussian posterior mean. Equation 1 is a local correction at each step, and it does not make the sampler draw from the exact posterior (Secs. 2 and 10).

Final paragraph (contribution, claims, evidence, and MRI). Our answers about the claims and the contribution are under the summary and point 8. The MRI experiments use four operators simulated on 256 × 256 face images with known coil sensitivities: one coil at 4-fold acceleration and eight coils at 2-, 4-, and 8-fold acceleration (App. B). We added them after the band's thresholds were fixed, as a test of the band (Sec. 6). With eight coils, the weak band grows from 0 to 0.241 as fewer lines are kept, and the cost of minimizing the residual grows from 0.016 to 0.209 (Table 3, Figure 2c). Figure 3 shows an 8-coil reconstruction at 8-fold acceleration, and Appendix J estimates the noise level from each MRI observation alone, to within 1.1 percent of the true level.

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

Related papers

Discussion (0)

Sign in to comment.