- 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
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 , 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 is 10). The failure on operators with a band extends to ΠGDM and partly to DPS, as reported above.
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 , defined in Sec. 2 as the smallest 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.
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 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.