acceptodds
Under review as a conference paper at ICLR 2027

Beyond Best-of-: Prior-Consistent DAPS for Ambiguous Nonlinear Inverse Problems

Abstract

Diffusion models have become strong priors for solving inverse problems. However, nonlinear inverse problems such as phase retrieval remain challenging when multiple measurement-consistent solutions are distinguishable only by the prior, we call such measurements ambiguous. Existing diffusion-based samplers rely on local approximations during reconstruction. For an ambiguous measurement, these approximations can allow measurement guidance to favor a measurement-consistent mode that is close to a prior-supported mode but itself not supported by the prior. Evaluation often masks this type of failure by reporting the best-of- runs, where the best is selected using ground truth. We introduce Prior-Consistent DAPS (PC-DAPS), a variant of Decoupled Annealing Posterior Sampling (DAPS) designed to address this failure case by maintaining multiple candidate solutions within a single sampling run. As annealing noise decreases, PC-DAPS progressively prunes and replaces weaker candidates according to their prior consistency, without any ground-truth access. PC-DAPS reduces the incidence of prior-inconsistent samples and improves nonlinear reconstruction quality while preserving sample quality on linear inverse problems.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
diffusion priors, posterior sampling, inverse problems

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.25 over 4 reviews, above 96% of the others.
Reviewer_uv9eOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce PC-DAPS, which aims to improve DAPS by pruning samples that could end up in the incorrect posterior modes during the sampling procedure, thus bypassing the need for best-of-k evaluation. The failure case, where standard DAPS can result in samples from the wrong mode, is clearly demonstrated using both toy examples and real data-driven priors. They propose two improvements during the sampling: (1) pruning intermediate particles that are unlikely under a surrogate prior density approximation using noised scores, and (2) combining the unimodal Gaussian posterior conditionals of each sample into a mixture of Gaussians. Experimental results demonstrate that PC-DAPS performs substantially better on multimodal inverse problems like phase retrieval, and samples the wrong mode much less often.

Critical strengths

The paper brings to light a significant limitation of DAPS, in which it often samples the wrong posterior mode, and also provides useful insights as to why and a solution. I have independently encountered similar issues using DAPS for posterior sampling, yet this problem has not really been discussed or analyzed in the literature despite the popularity of DAPS, making the paper's focus a valuable contribution.

The proposed improvements are intuitive, efficient, and justifiable under the Gaussian approximation.

For the same amount of computation and sampled particles, PC-DAPS outperforms even the best-of-k DAPS under image-based metrics, making it practically useful when generating many samples.

Critical weaknesses

SMC-based methods (like TDS) also perform resampling, which also weight and prune particles depending on their trajectory, but their performance on phase retrieval is not shown. What is the advantage of PC-DAPS over existing SMC approaches? Does PC-DAPS give more diversity and less particle collapse under the same number of samples?

Questions for authors

DAPS also has a diffusion-score alternative where the Gaussian conditional is replaced with the learned score queried at low noise. The samples of p(x0∣xt,y)p(x_0 \mid x_t, y) should be more prior-consistent due to not relying on Gaussian p(x0)p(x_0) approximation, making upside down / inverted images much less likely, though with more computation and maybe less measurement consistency. How does PC-DAPS compare?

How do the surrogate prior-based weights compare to the true SMC weights along the DAPS t→p(xt∣y)t \to p(x_t \mid y) annealing path?

What happens when the measurement is OOD, i.e. where the ground truth already fails the calibration check? Does PC-DAPS give worse reconstructions than standard DAPS here?

3 comments
AuthorsOpenReview ↗

Response to uv9e (1 of 2)

We thank the reviewer for the constructive feedback and for recognizing (i) the significance of the wrong-mode sampling issue in DAPS, which the reviewer has also independently observed, (ii) the intuitive, efficient, and theoretically grounded nature of our improvements, and (iii) the practical gains of PC-DAPS at equal compute. We address the remaining concerns below. Our citations appear above as a single public comment, as the same numbers are used across comments.

We will answer both (W1) and (Q1) using the following table:

MethodPSNR (↑\uparrow)SSIM (↑\uparrow)LPIPS (↓\downarrow)Flipped (↓\downarrow)Failed (↓\downarrow)In-bag LPIPSTime / image
PC-DAPS30.060.8340.1511.5%1.0%0.09930 s
TDS, 4 particles18.800.5450.40320.0%12.0%0.000116 s
DAPS, Gaussian step25.130.6680.28314.0%6.0%0.10028 s
DAPS, diffusion-score step26.000.6740.27912.5%5.0%0.101707 s

(W1) SMC-based methods (like TDS) also perform resampling, which also weight and prune particles depending on their trajectory, but their performance on phase retrieval is not shown. What is the advantage of PC-DAPS over existing SMC approaches? Does PC-DAPS give more diversity and less particle collapse under the same number of samples?

Because PC-DAPS uses a generative prior, we focus on generative (specifically diffusion) SMC-based methods, for example [1, 3, 9]. In general, the main difference we see between the three SMC-based methods and PC-DAPS is that the SMC-methods typically use measurement-based importance weights, whereas PC-DAPS uses prior-consistency scores for pruning. However, of these three methods with which we are most familiar, FPS-SMC [3] and MCGDiff [1] are constructed to solve linear inverse problems. Because PC-DAPS is particularly interesting for nonlinear inverse problems, we compare against TDS, which supports nonlinear measurement models.

As you note, TDS [9] doesn’t report phase-retrieval results; thus, we implement our own phase retrieval twisting function and use their official code with the default settings, 4 particles (to match the PC-DAPS bag size), and the same diffusion model. We obtain an average PSNR of 18.80 dB and observe that 20% of its samples are flipped or failed (cf. 1.5% from PC-DAPS). We observe the cause of this TDS failure to be weight degeneracy: with 4 particles and 442k measurements, the median empirical effective sample size is 1.0 over the first 40% of the reverse process, so one particle takes all the weight from the start. On all 100 trials, only one of the 4 initial particles has descendants at the end, and the 4 outputs are near-copies (in-bag pairwise LPIPS 0.000). TDS thus returns one reconstruction per measurement; when it is flipped, all 4 samples are flipped. Also, PC-DAPS is 3.9 times faster than TDS, since it does not use backpropagation during sampling. All the results are presented in the table above.

(Q1) DAPS also has a diffusion-score alternative where the Gaussian conditional is replaced with the learned score queried at low noise. The samples of p(x0∣xt,y)p(x_0 \mid x_t, y) should be more prior-consistent due to not relying on Gaussian p(x0)p(x_0) approximation, making upside down / inverted images much less likely, though with more computation and maybe less measurement consistency. How does PC-DAPS compare?

We ran the diffusion-score option in DAPS’s official code base from the same initialization as the Gaussian step, and find the quality gain to be small despite the large increase in cost. Compared to DAPS-1K, the proportion of flipped samples in FFHQ phase retrieval decreases from 14.0% to 12.5% (cf. 1.5% from PC-DAPS) and metric gains are small (e.g. +0.87 dB for PSNR), yet the wall clock sampling time increases from 28s to 707s. All the results are presented in the table above.

AuthorsOpenReview ↗

Response to uv9e (2 of 2)

(Q2) How do the surrogate prior-based weights compare to the true SMC weights along the DAPS t→p(xt∣y)t \to p(x_t \mid y) annealing path?

We are very interested in the answer to this question. We find it difficult to compare the two, primarily because computing accurate posterior weights is difficult when solving nonlinear inverse problems. For example, in IEnSF [11], the authors aim to solve nonlinear inverse problems with Gaussian mixture priors (appearing as the analysis step in data assimilation). The authors state at the beginning of Section 3.2 that, even in their GMM-only prior setting, the posterior mixture weights are intractable when A\mathcal{A} is nonlinear and thus must be approximated (when A\mathcal{A} is linear the weights are available by Gaussian conjugacy).

In our setting, our understanding is that, to know the analog of SMC weights along the DAPS path, we would require knowledge of the Langevin proposal p(y∣x0)p(x0∣xt)p(y \mid x_0)p(x_0 \mid x_t) and/or its integral over x0x_0, the denoising likelihood p(y∣xt)p(y \mid x_t). Unfortunately, it’s unclear to us how to accurately estimate either in high dimensions, even offline, since we don’t know p(x0∣xt)p(x_0 \mid x_t). However, being able to tractably estimate these quantities in high dimensions would make for a great benchmark for existing twisting and tilting methods, and would resolve the final paragraph of our discussion section; thus, we have great interest in accurate techniques for computing these weights. We see the intractability of these objects as the primary motivation for using methods like DPS and DAPS.

Finally, although we can compare our prior-consistency scores to the SMC weights (given access to the latter), we want to flag that since SMC weights in generative approaches are typically importance weights, they use likelihood information, and so the scores and weights are likely quite dissimilar.

(Q3) What happens when the measurement is OOD, i.e. where the ground truth already fails the calibration check? Does PC-DAPS give worse reconstructions than standard DAPS here?

To measure OOD behavior, we ran PC-DAPS with the FFHQ score conditioned on the phase retrieval measurements of both (i) FFHQ validation images (in-distribution) and (ii) ImageNet images (OOD). Results reported in the table below are validated across 50 measurements.

On the ImageNet examples, we observe that PC-DAPS obtains better performance than DAPS: the average PSNR of the PC-DAPS samples is marginally higher, and the orientation of the outputted samples is consistent with the ground truth’s orientation more often.

yy datasetGT fails checkPSNR (↑\uparrow)Correct orientation (↑\uparrow)
PC-DAPSDAPSPC-DAPSDAPS
FFHQ0/5029.6525.0294%78%
ImageNet0/5016.4614.7346%32%
Reviewer_uv9eOpenReview ↗

Response to Authors

I appreciate the author's thorough response, and my concerns are resolved. As this is a direct improvement to a state-of-the-art posterior sampling method, I'll raise my score to 4.

Reviewer_s8hAOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces PC-DAPS to reduce measurement-consistent but prior-inconsistent reconstructions in ambiguous inverse problems. It maintains multiple particles, constructs a Gaussian mixture approximation of the denoising distribution, and applies a one-shot pruning and replacement procedure using denoising-based consistency scores. Experiments show substantial improvements in phase retrieval on FFHQ and more modest gains on ImageNet.

Critical strengths
  1. The paper addresses a meaningful evaluation problem: ground-truth-based best-of-(k) selection can conceal unreliable individual reconstructions. Appendix Table 4 provides a useful matched mean-of-(k) comparison.
  2. The improvements on FFHQ are substantial. Table 2 reports an increase in faithful reconstructions from 79.5% to 97.5%, with failed reconstructions decreasing from 6.0% to 1.0%.
  3. The toy examples provide useful intuition, and the paper acknowledges limitations concerning ImageNet and reference-based reconstruction metrics.
Critical weaknesses
  1. The prior-consistency score lacks sufficient validation. Equation (9) measures denoising reconstruction error, but the paper does not establish when this quantity reliably ranks prior support or probability. A plausible sample can change substantially under denoising at larger noise scales. Calibration using within-cluster score variation controls a threshold but does not establish discrimination between valid and invalid modes. This is central because an incorrect ranking can permanently remove useful trajectories.
  2. The conditional interpretation of the GMM is unclear. Equations (10)–(12) pool estimates from particles with different (x_t^{(k)}), yet describe the resulting distribution as an approximation to (p(x_0\mid x_t)). The derivation does not explain why this shared mixture approximates each particle’s conditional distribution or what distribution the resulting coupled updates target. The method is empirically motivated, but its probabilistic interpretation needs qualification.
  3. Removing hallucinated modes does not establish preservation of the posterior. Figure 2 explicitly shows that PC-DAPS also misses a valid minority mode. Under the parameters in Appendix A.2, this mode carries approximately (1/(1+e^2)=11.9%) posterior probability. The authors acknowledge the omission, but the qualitative within-bag diversity in Figure 3 does not establish recovery of distinct valid modes or their probabilities.
Questions for authors

see weaknesses

3 comments
AuthorsOpenReview ↗

Response to s8hA (1 of 2)

We thank the reviewer for the constructive feedback and for recognizing the importance of the problem we raise and the intuition provided by our toy examples. We address the remaining concerns below. Our citations appear above as a single public comment, as the same numbers are used across comments.

(W1) The prior-consistency score lacks sufficient validation. Equation (9) measures denoising reconstruction error, but the paper does not establish when this quantity reliably ranks prior support or probability. A plausible sample can change substantially under denoising at larger noise scales. Calibration using within-cluster score variation controls a threshold but does not establish discrimination between valid and invalid modes. This is central because an incorrect ranking can permanently remove useful trajectories.

For theoretical grounding, we claim that Eq. 9 measures distance to the data manifold, which we justify as follows. First, note that our pruning step only runs at small noise levels σt\sigma_t. It’s known that, at sufficiently small noise levels, the learned score is orthogonal to the manifold [7]. Thus, while a sample on the manifold changes only by denoising along the manifold, giving a squared norm error of about σ2r\sigma^2 r (where rr is the intrinsic dimension), a sample with distance δ\delta off the manifold gives σ2r+δ2\sigma^2r + \delta^2. Because we only rank candidates for the same measurement, the shared σ2r\sigma^2 r term cancels (assuming on-manifold samples denoise by relatively similar amounts) and the ranking is dominated by δ\delta, which (as explained above) is what the score is being used to estimate. Eq. 9 can alternatively be interpreted as a truncated and reweighted form of the diffusion ELBO, so lower values loosely track higher likelihood under the prior. Empirically, we have checked and confirm that no valid samples are pruned in our experiments.

(W3) Removing hallucinated modes does not establish preservation of the posterior. Figure 2 explicitly shows that PC-DAPS also misses a valid minority mode. Under the parameters in Appendix A.2, this mode carries approximately 1/(1+e2)=11.9%1/(1+e^2)=11.9\% posterior probability. The authors acknowledge the omission, but the qualitative within-bag diversity in Figure 3 does not establish recovery of distinct valid modes or their probabilities.

We acknowledge that our method might not cover all posterior modes, but claim that we only remove bad modes from the DAPS posterior. We find that existing methods do some combination of (i) sampling from modes that shouldn’t be there and (ii) failing to sample from modes that should be there. Our paper’s goal is to design techniques which directly address issue (i) without worsening issue (ii).

To justify that our method does not worsen issue (ii), we aim to demonstrate that our pruning does not remove any valid mode that DAPS samples from, implying that we only remove illegitimate uncertainty arising from issue (i). To showcase this, we modify our toy example in Figure 2 by adding an additional mode that is both prior- and likelihood-consistent, meaning DAPS samples from it, but has minority weight. As shown in the table below, PC-DAPS samples from all posterior-consistent modes that DAPS samples from, but removes prior-inconsistent modes. Additionally, observe that our relative weighting of the first and second columns is more accurate than the weighting of DAPS, which overweights the second column.

MethodMajority(New) minorityPrior inconsistent at +1.1+1.1Prior consistent at +1+1
Exact87.6%10.9%0%1.5%
DAPS76.2%12.1%11.7%0%
PC-DAPS88.4%11.6%0%0%

To demonstrate diversity on a real image task, we utilize image quality as a proxy and demonstrate diversity of LPIPS within a bag (using pairwise LPIPS distance) for tasks which we reason should exhibit posterior diversity: box inpainting and 16$\times$ super-resolution. PC-DAPS exhibits nearly identical diversity as DAPS. Both tasks are on FFHQ. If you have a specific diversity or mode coverage metric in mind, please let us know.

TaskMethodIn-run/bag LPIPS ↑\uparrow
Inpaint (190×190190\times190 box)DAPS0.215
PC-DAPS0.217
Super-resolution 16×16\timesDAPS0.191
PC-DAPS0.188
AuthorsOpenReview ↗

Response to s8hA (2 of 2)

(W2) The conditional interpretation of the GMM is unclear...needs qualification.

We think that this is a great and timely question. We agree that Eqs. (10)–(12) are dissimilar in distribution to the true denoising distribution p(x0∣xt)p(x_0 \mid x_t), and we are happy to both qualify this in Section 4.2 and discuss it at greater length in the Appendix. We claim that, although we’d like for our approximation to be accurate in a distributional sense, in practice this is not tractable, so it is common to instead select the denoising distribution to maximize accuracy of a downstream task.

For example, in tilting methods (such as DPS), the goal is to guide the diffusion model with ∇xtlog⁡p(y∣xt)\nabla_{x_t} \log p(y \mid x_t), which can be written in integral form but does not typically simplify algebraically for the true p(x0∣xt)p(x_0 \mid x_t), even if it were known. For example, if A\mathcal{A} is nonlinear, then the integral only simplifies if p(x0∣xt)p(x_0 \mid x_t) is discrete (e.g., if p(x0)p(x_0) is discrete). Thus, the approximation’s primary goal isn’t to capture p(x0∣xt)p(x_0 \mid x_t) well; it’s to approximate p(x0∣xt)p(x_0 \mid x_t) such that the score of the resulting p(y∣xt)p(y \mid x_t) approximation is (most importantly) computationally tractable, with a secondary goal of accuracy.

In our case, we want an approximation q(x0∣xt)≈p(x0∣xt)q(x_0 \mid x_t) \approx p(x_0 \mid x_t) such that sampling from p(y∣x0)q(x0∣xt)p(y \mid x_0)q(x_0 \mid x_t) via ULA behaves as much like sampling from p(y∣x0)p(x0∣xt)p(y \mid x_0)p(x_0 \mid x_t) as possible. Existing methods which are popular for image generation (e.g., DAPS, Π\PiGDM [6]) typically use a unimodal approximation of a fixed (Gaussian) class with an isotropic and data-independent covariance. Thus, to make an improvement on the existing approximations, we look for a similarly tractable approximation that captures (i) the potential multimodality of p(x0∣xt)p(x_0 \mid x_t) and (ii) directions of variability in the data.

The relevant difference between the DAPS and PC-DAPS approximations is how the two choices affect the Langevin drift. DAPS’s Gaussian contributes rt−2(x^0(xt)−x0)r_t^{-2}(\hat x_0(x_t)-x_0), pulling every iterate toward the ODE endpoint where ULA was initialized. Thus, because no other terms in the Langevin update incorporate prior information, if a Langevin chain is walking towards two measurement-consistent wells of the likelihood potential, of which only one is prior-consistent, the DAPS Langevin would not be able to discern which is which. Instead, the DAPS Langevin will favor whichever well is closer to the initialization x^0(xt)\hat x_0(x_t).

A first improvement would be to at least (i) encourage movement in more prior-plausible directions and to discourage movement in less prior-plausible directions, for example by using an anisotropic covariance, and (ii) encourage movement in each pixel at a scale consistent with the expected variability in the data, for example by replacing the data-independent scale rt2r_t^2 with a data-based scaling. We claim that replacing the isotropic covariance rt2r_t^2 with the covariance in each pixel over the elements of the bag or cluster, as is done in PC-DAPS, is an improvement along both axes.

A second improvement would be that, in cases where moving away from x^0(xt)\hat x_0(x_t) might improve measurement and/or posterior consistency, we’d want our Langevin update to at least allow for the iterate to travel away from its initialization. A loftier goal would be for the Langevin updates to accurately detect which of the two likelihood-consistent wells in the above example is prior-consistent. This motivates the pre-pruning Gaussian mixture: if there is a particle near one of the two likelihood wells, that particle was an unconditional prior sample denoised from an xtx_t at a high noise level, and thus is likely in a prior-consistent region. Hence, we claim one should favor that well. Post-pruning, the remaining components represent only modes that the bag has collectively judged prior-consistent. Thus, the mixture drift can only favor wells that survived the prior-consistency check in Eq. (9). Even if the Langevin update does not leave the identified mode, the Langevin dynamics now discourages iterates from leaving the prior mode identified by the cluster rather than from the Euclidean ball around the ODE endpoint, which may have significant off-prior volume.

To explicitly answer the last part of your question (what the coupled updates target), qq is built from and shared across the bag, rather than from xt(k)x_t^{(k)} individually. However, xt(k)x_t^{(k)} enters through (i) the particle’s Langevin initialization and (ii) which component dominates the weights qm(x0)q_m(x_0). When these weights are near one-hot, as is typical in high dimensions and/or at initialization, a particle in cluster mm targets N(x0;μm,Σm)p(y∣x0)\mathcal{N}(x_0;\mu_m,\Sigma_m) p(y \mid x_0), while all DAPS particles target N(x0;x^0(xt),rt2I)p(y∣x0)\mathcal{N}(x_0;\hat x_0(x_t),r_t^2 I) p(y \mid x_0).

Dear Reviewer, As the discussion period comes to an end, we would like to ask whether our responses have addressed your initial concerns and if you have any remaining questions or concerns.

Reviewer_xS32OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Prior-Consistent DAPS (PC-DAPS), a particle-based extension of DAPS for inverse problems with ambiguous, multi-modal posteriors. The motivation is that a reconstruction can match the measurements and still be implausible under the learned prior. Standard best-of-k evaluation, which keeps the sample closest to the ground truth, can hide this failure. PC-DAPS runs multiple interacting particles and builds a Gaussian-mixture approximation from their denoised estimates. It then uses a denoising-based prior-consistency score (Eq. 9) to prune and replace implausible candidates without access to the ground truth. Experiments cover seven linear and three nonlinear tasks on FFHQ and ImageNet. They report standard reconstruction metrics and how often each failure type occurs. The clearest gain is on FFHQ phase retrieval: faithful reconstructions rise from 79.5% to 97.5%, and upside-down or colour-inverted outputs fall from 14.5% to 1.5%.

Critical strengths
  1. It provides the diagnoses of the failures hidden by best-of-k evaluation. It explains how DAPS can become trapped in reconstructions that match the measurements but are implausible under the image prior.
  2. PC-DAPS maintains multiple interacting particles, constructs a Gaussian-mixture approximation from their denoised estimates, and uses a denoising-based consistency score to prune and replace problematic candidates without ground-truth access.
  3. The experimental study is comprehensive. Experiments cover seven linear and three nonlinear tasks on FFHQ and ImageNet. The evaluation includes average performance across samples and explicit failure frequencies, alongside conventional reconstruction metrics.
Critical weaknesses
  1. Posterior fidelity is not validated, and pruning may discard legitimate uncertainty. Pruning by prior consistency risks removing valid but low-probability modes, which would make a posterior sampler behave more like a mode-seeking method. Figure 2 itself shows that a valid minority mode is under-sampled. The method is presented as posterior sampling for ambiguous problems, so this point is central. The paper does not show that sample frequencies match posterior mass, for example on a synthetic problem with a tractable posterior. Without that, the lower failure rates may have been bought at the cost of calibrated uncertainty.

  2. Gains on standard metrics are modest, and their source is not isolated. In Table 1, PC-DAPS improves only marginally over DAPS on most tasks. Table 3 studies ODE scheduling and removing the final Langevin update. However, there is no real-image ablation that separates pruning, the Gaussian-mixture approximation and their combination. It is also unclear whether the gains hold at matched computation, since PC-DAPS runs several interacting particles. A particularly informative baseline would be independent DAPS runs with the same total compute, ranked by the same prior-consistency score. If that baseline performs comparably, the contribution comes down mainly to the selection criterion rather than the particle interaction.

Questions for authors
  1. On the real-image tasks, could you report results for (a) pruning only, (b) the Gaussian-mixture approximation only, (c) both combined, and (d) independent DAPS runs selected with the same prior-consistency score? Please run all of them at matched total computation (e.g. number of function evaluations or wall-clock time). Please also confirm whether the implementation changes in PC-DAPS, such as the ODE scheduling, were also applied to the DAPS baseline.

  2. How reliably does the score in Eq. (9) tell prior-inconsistent samples apart from valid but uncommon reconstructions? Evidence could include the score distributions of the two groups in phase retrieval, the rate at which valid samples are wrongly pruned, or results on a synthetic problem with a known posterior.

3 comments
AuthorsOpenReview ↗

Response to xS32 (1 of 2)

We thank the reviewer for the constructive feedback and for recognizing (i) our diagnosis of how DAPS can become trapped in measurement-consistent but prior-implausible reconstructions that best-of-kk evaluation hides, (ii) the design of PC-DAPS, which detects and replaces such failures without ground-truth access, and (iii) the comprehensiveness of our experiments. We address the remaining concerns below. Our citations appear above as a single public comment, as the same numbers are used across comments.

(W1) Posterior fidelity is not validated, and pruning may discard legitimate uncertainty. Pruning by prior consistency risks removing valid but low-probability modes, which would make a posterior sampler behave more like a mode-seeking method. Figure 2 itself shows that a valid minority mode is under-sampled. The method is presented as posterior sampling for ambiguous problems, so this point is central. The paper does not show that sample frequencies match posterior mass, for example on a synthetic problem with a tractable posterior. Without that, the lower failure rates may have been bought at the cost of calibrated uncertainty.

We acknowledge that our method might not cover all posterior modes, but do not find that it reduces legitimate uncertainty. We find that existing methods do some combination of (i) sampling from modes that shouldn’t be there and (ii) failing to sample from modes that should be there. Our paper’s goal is to design techniques which directly address issue (i) without worsening issue (ii).

To justify that our method does not discard legitimate uncertainty, we demonstrate that our pruning does not remove valid modes that DAPS samples from, implying that we only remove illegitimate uncertainty arising from issue (i). We modify our toy example in Figure 2 by adding an additional mode that is both prior- and likelihood-consistent, but has minority weight. As shown in the table below, PC-DAPS samples from all posterior-consistent modes that DAPS samples from, including the new minority mode, but removes prior-inconsistent modes. Additionally, observe that our relative weighting of the first and second columns is more accurate than the weighting of DAPS, which overweights the second column.

MethodMajority(New) minorityPrior inconsistent at +1.1+1.1Prior consistent at +1+1
Exact87.6%10.9%0%1.5%
DAPS76.2%12.1%11.7%0%
PC-DAPS88.4%11.6%0%0%
AuthorsOpenReview ↗

Response to xS32 (2 of 2)

(W2) Gains on standard metrics are modest, and their source is not isolated. In Table 1, PC-DAPS improves only marginally over DAPS on most tasks. Table 3 studies ODE scheduling and removing the final Langevin update. However, there is no real-image ablation that separates pruning, the Gaussian-mixture approximation and their combination. It is also unclear whether the gains hold at matched computation, since PC-DAPS runs several interacting particles. A particularly informative baseline would be independent DAPS runs with the same total compute, ranked by the same prior-consistency score. If that baseline performs comparably, the contribution comes down mainly to the selection criterion rather than the particle interaction.

(Q1) On the real-image tasks, could you report results for (a) pruning only, (b) the Gaussian-mixture approximation only, (c) both combined, and (d) independent DAPS runs selected with the same prior-consistency score? Please run all of them at matched total computation (e.g. number of function evaluations or wall-clock time). Please also confirm whether the implementation changes in PC-DAPS, such as the ODE scheduling, were also applied to the DAPS baseline.

As requested, we ran an ablation on FFHQ phase retrieval (100 images, k=4k=4) and report mean-of-kk for all metrics in the table below. The reported DAPS baseline numbers in our paper (first row) are the original DAPS-1K method without modification and lack our implementation changes, as the original method has better performance in metrics than DAPS-1K with our modifications (second row).

The dominant cost for all methods in the table is from NFEs from denoising, so we list NFEs instead of wall clock time. Note it is difficult to exactly match the total number of NFEs across rows, but that NFEs are similar for all rows but one: the DAPS-1K row uses 1k NFEs per sample (200 steps with 5 NFEs each), while all other rows use at most 735 NFEs per sample. 732 of the 735 come from our modified ODE (200 steps with our 1→51\to5 ODE schedule) and the remaining 3 come from the ∣S∣=3|\mathcal{S}|=3 NFEs per sample in the single iteration where pruning/selection happens. In the table, (ref) is DAPS-1K + ODE schedule + Langevin drop

MethodPSNR (↑\uparrow)SSIM (↑\uparrow)LPIPS (↓\downarrow)Faithful (↑\uparrow)Upside-down/color-inverted (↓\downarrow)NFEs# samples
DAPS-1K25.110.6660.283318 (79.5%)58 (14.5%)1000400
(ref)25.020.6610.284315 (78.8%)64 (16.0%)732400
(ref) + mixture25.990.7030.241337 (84.3%)44 (11.0%)732400
(ref) + pruning29.220.8010.160380 (95.0%)12 (3.0%)735400
PC-DAPS-1K30.090.8340.152390 (97.5%)6 (1.5%)735400
(ref) + selection28.080.7680.191275 (90.2%)26 (8.5%)735$^*$305

We observe that our pruning technique produces the majority of PC-DAPS’s reported improvement (e.g., adding pruning to the reference raises PSNR by 4.2 dB). The Gaussian mixture alone yields only modest metric gains (e.g., +0.97 dB above ref in PSNR) but still reduces the proportion of upside-down/color-inverted samples; thus, we view the GMM as a complementary improvement.

Using prior-consistency selection (final row) to discard bad outputs from the reference gives improved metrics among the retained samples (e.g., +3.06 dB above ref in PSNR). However, this increases the number of NFEs per retained sample to 735∗400/305≈964735 * 400 / 305 \approx 964 (hence the asterisk) only to obtain worse performance in all columns than both the reference w/ pruning and the full PC-DAPS method.

(Q2) How reliably does the score in Eq. (9) tell prior-inconsistent samples apart from valid but uncommon reconstructions? Evidence could include the score distributions of the two groups in phase retrieval, the rate at which valid samples are wrongly pruned, or results on a synthetic problem with a known posterior.

From the table provided in response to (W2) and (Q1) above, when we turn pruning off (mixture only), the flipped samples increase from 1.5% to 11.0% and the faithful samples decrease from 97.5% to 84.3%. Relative to the reference sampler, the mixture alone reduces flipped samples from 16.0% to 11.0%, and relative to reference w/ pruning the mixture further reduces them from 3.0% to 1.5%.

For the prior score in Eq. (9), the score separates invalid from valid samples with an AUROC of 0.95. Relative to valid (upright) reconstructions, the mean prior score is 69% worse for color-inverted samples and 21% worse for rotated samples, making rotation the harder ambiguity. Overall, we find that no valid reconstructions are pruned.

Dear Reviewer, As the discussion period comes to an end, we would like to ask whether our responses have addressed your initial concerns and if you have any remaining questions or concerns.

Reviewer_KXbqOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3(+2)
Read review
Paper summary

This paper looks at diffusion-based solvers for nonlinear inverse problems where several solutions fit the measurement, but only some are plausible under the prior. Fourier phase retrieval is the main example, where solvers can return flipped or color-inverted faces. The authors point out that the usual practice of reporting the best of several runs, chosen using the ground truth, hides this problem. They propose PC-DAPS, a variant of DAPS that runs a small set of particles together. The particles share a Gaussian mixture approximation of the denoising distribution, and at one point during sampling, groups of particles that look less consistent with the prior are replaced by copies of the better ones.

The method is tested on FFHQ and ImageNet on seven linear tasks and three nonlinear ones (phase retrieval, nonlinear deblurring and HDR), against DAPS, DPS and several other diffusion baselines. The authors also count how often phase retrieval outputs come out correct, flipped or failed, and report fewer flipped outputs than independent DAPS runs, especially on FFHQ. Ablations cover the number of ODE steps, annealing steps and particles, as well as runtime.

Critical strengths

The paper points to a real weakness in how diffusion solvers are evaluated on phase retrieval. Reporting the best of several runs, chosen using the ground truth, hides outputs that fit the measurement but are implausible. The breakdown of outputs into correct, flipped and failed (Table 2) and the comparison that averages over samples for every method (Table 4) make this visible.

On FFHQ phase retrieval, the method reduces these implausible reconstructions without using the ground truth, at a runtime comparable to several independent DAPS runs. This is the main evidence that the idea works, though it is limited to one task and the gain is much smaller on ImageNet.

Critical weaknesses

I have noticed several issues in the paper. I will list some of the major ones here.

  • Each particle starts at its own mixture component, and in such a high-dimensional space the other components are far away, so the mixture weights should be essentially one-hot. The mixture then cannot pull particles toward other modes as the toy example suggests, and the authors should show the mixture weights during sampling. In addition, as the paper itself notes through Jensen’s inequality, pruning cannot trigger on linear problems, and most tasks in Table 1 are linear.

  • The method description leans heavily on DAPS and defers several details either to the supplementary material or to other papers. For example, Algorithm 1 should be described in the main paper. Some parameters, like Var_B, are never defined. In addition, a few steps of Algorithm 1 do not match the text.

  • The paper claims the method preserves diversity (Figures 11 and 12), but no diversity metric is reported. After pruning, all particles in a cluster are pulled toward the same cluster mean, which should reduce diversity. For tasks with broad posteriors, such as 16× super-resolution and box inpainting, the authors should report a diversity measure across samples.

  • Most of the inverse problems are not described well. The paper should be self-contained, and it is not reasonable to defer all the details of the forward operators to another paper. Also, the paper mostly uses “phase retrieval” when it really means “Fourier phase retrieval”. I suggest the authors correct this terminology. Phase retrieval is a broad class of problems that covers many measurement models and often complex-valued signals, but here only the Fourier magnitude of a real image is measured. Also, what exactly is nonlinear deblurring?

  • Visual Results for competing methods are only shown in Figures 1 and 3, both for phase retrieval. Even in the supplementary material, the authors only show results from their own method. Why are there no visual comparisons with the baselines for the other tasks?

  • Why are the results for the DAPS baseline below the published numbers? Also, some of the results look off. For example, on ImageNet phase retrieval DPnP has a higher SSIM than PC-DAPS (0.487 vs 0.469) despite a PSNR about 5 dB lower (16.74 vs 21.75 dB).

Questions for authors
  • What part of the method prevents flipped outputs? If it is the pruning step, please show Table 2 with pruning turned off, and report how well the prior score in Eq. 9 tells correct clusters apart from flipped ones.
  • The mixture uses the variance of the denoised estimates across particles as its covariance. Why is this the right choice for the denoising distribution?
  • How does this covariance compare to the DAPS covariance schedule across noise levels, and is a lower bound used to keep it from going to zero?
  • On ImageNet, a flipped image can be just as plausible under the prior as the upright one. Do the flipped outputs actually get a worse prior score, or are they valid posterior samples?
  • In Table 1, the baselines are reported as the best of four runs while PC-DAPS is reported as the average of its four samples. Why not use the same protocol for all methods? The comparison in the supplementary that averages for every method should be moved to the main paper.
  • For ImageNet phase retrieval, the standard deviations in Table 1 seem too small given how many outputs Table 2 labels as flipped or failed. Could the authors report the metrics separately for each outcome type?
7 comments
AuthorsOpenReview ↗

Response to KXbq (1 of 3)

We thank the reviewer for the detailed feedback and for recognizing the evaluation weakness we expose, namely that ground-truth-based best-of-k selection hides measurement-consistent but implausible outputs, and that PC-DAPS reduces such failures without ground-truth access at runtime. We address each concern below, labeling the weaknesses W1-W6 and the questions Q1-Q6 in the order they appear in the review. Our citations appear above as a single public comment, as the same numbers are used across comments.

(W1, mixture weights) We take “mixture weights” to mean the score weights qm(x0)q_m(x_0) in Equation (13), not the mixture weights πm\pi_m in Equations (10)-(12). We agree that qm(x0)q_m(x_0) can be one-hot: for example, PC-DAPS on ImageNet HDR and 4$\times$ super-resolution produce one-hot qm(x0)q_m(x_0) on over 99.99% of all Langevin iterations. However, we observe particles switching GMM components in about 30% of chains (29% for HDR, 30% for super-resolution), and we find that switches happen exclusively at high and moderate noise levels (σt≥2\sigma_t \geq 2).

We acknowledge that the toy example shows smoother weight variation than the high dimensions, where we observe abrupt switches (e.g., qmq_m jumps from one-hot for component mm to one-hot for component m′m' with ≤5\leq 5 Langevin steps). We will revise lines 784-785 and Appendix A.4 accordingly, and add plots of qm(x0)q_m(x_0) during sampling to the Appendix.

The mixture also has practical benefits: as described in more detail in response to Q1, we compare a reference (DAPS-1K with our ODE and Langevin settings) with vs. without the Gaussian mixture on FFHQ phase retrieval. Using the mixture (without pruning) increases the proportion of faithful samples from 78.8% to 84.3% and decreases the proportion of flipped samples from 16.0% to 11.0%.

(W1, linear tasks) We include all the linear tasks in Table 1 to compare PC-DAPS against baselines on all tasks which the original DAPS paper investigates, to show that PC-DAPS is not restricted to nonlinear inverse problems even though it’s true that the pruning only triggers on nonlinear problems.

(W2, DAPS dependence) Although our algorithm is presented as a modification of DAPS, which we chose for its strong performance, we emphasize that the prior-consistency of other existing algorithms can also be improved using our technique(s). For example, one can obtain substantial improvements by modifying PnP-DM, shown below on FFHQ phase retrieval with mean-of-kk evaluation:

MethodPSNR (↑\uparrow)SSIM (↑\uparrow)LPIPS (↓\downarrow)Faithful (↑\uparrow)
PnP-DM17.68±5.5017.68_{\pm 5.50}0.458±0.1910.458_{\pm 0.191}0.523±0.1490.523_{\pm 0.149}54.0%
PnP-DM w/ pruning25.45±4.9125.45_{\pm 4.91}0.749±0.1200.749_{\pm 0. 120}0.274±0.1150.274_{\pm 0.115}79.8%

We show that (i) existing algorithms have issues that are masked by best-of-kk evaluation but (ii) can be modified in a relatively simple manner to decrease the prevalence of prior-inconsistent samples without increasing computational cost. We see the modification as a desirable property: rather than having to familiarize yourself with an entirely new algorithm, its quirks, its tuning, etc., you can take an algorithm you already like and patch it using the techniques we describe.

(W2, Algorithm 1 placement) Algorithm 1 is described in the main paper in Section 4. Our aim in Section 4 is to emphasize (i) our contributions and (ii) the most critical parts of our method. Thus, we defer implementation details of DAPS and subroutines from the literature (such as clustering or the Bayesian information criterion) to the Appendix, as we aren’t claiming these implementation details as novel. The exception to this is that we describe our novel ODE schedule and the choice to omit the last Langevin step in the Appendix, but note that these choices are (i) best motivated by their accompanying performance and wall clock time ablations in the Appendix and (ii) not essential for our method to improve over DAPS in metrics (e.g., the ODE schedule is for runtime). Finally, we are happy to define any notation which isn’t clear, such as by explicitly defining VarB\mathrm{Var}_{\mathcal B} as the variance of the elements in the set B\mathcal{B}.

(W2, Algorithm 1 typesetting) Thank you for catching these. The main mismatch is that Algorithm 1 has Langevin starting from the Tweedie estimate x~0(k)\tilde x_0^{(k)}, while the text starts it from the ODE endpoint x^0(xt(k))\hat x_0(x_t^{(k)}), as in DAPS. These coincide at high noise, where we use a single Euler step (which returns exactly Dθ(xt;σt)D_\theta(x_t;\sigma_t)), but differ at low noise, where we use five Euler steps. We will fix Algorithm 1, and also make the final-step indexing, replacement rule, and pruning test notation consistent.

AuthorsOpenReview ↗

Response to KXbq (2 of 3)

(W3) To our knowledge, developing diversity metrics for generative modeling is a significant open problem, as the true prior and posterior are both unknown. To use image quality as a proxy, we report average pairwise LPIPS of DAPS and PC-DAPS FFHQ outputs to demonstrate that our outputs (i) are not kk identical copies of one sample and (ii) have similar diversity to the DAPS outputs. If you have a specific diversity metric in mind, please let us know.

TaskMethodIn-run/bag LPIPS ↑\uparrow
Inpaint (190×190190\times190 box)DAPS0.215
PC-DAPS0.217
Super-resolution 16×16\timesDAPS0.191
PC-DAPS0.188

(W4) We are happy to include an explicit inverse problem definitions in the Appendix. We, like the DAPS paper, use the inverse problem setups of the DPS paper (i.e., DAPS defers the details to DPS). However, we acknowledge that the deferral is inconvenient (e.g., if a reader hasn’t read the DPS or DAPS papers) and so will add the definitions of all inverse problems to the Appendix.

We explicitly call the task “Fourier phase retrieval” when the problem is introduced at the beginning of the second paragraph in the introduction, and refer to it as “phase retrieval” in our tables for (i) brevity and (ii) consistency with the DPS and DAPS baselines. We acknowledge that phase retrieval can mean a broader class of inverse problems, but note that our usage of the term is consistent with the seminal work that we cite when first defining the problem [4].

For nonlinear deblurring, like DPS and DAPS we use the learned blur operator of Tran et al. [8] (bkse) as the forward model. This operator is a neural network Fϕ(x,k)\mathcal{F}_\phi(x, k) trained on blurry and sharp image pairs from the GOPRO dataset [5], in which real motion blur is created by averaging consecutive high-speed video frames under a nonlinear camera response function. Given a sharp image xx and a latent kernel code kk, the network outputs the corresponding blurry image, so unlike convolutional blur the forward operator is nonlinear in xx.

(W5) We are happy to include additional visual baseline comparisons in the Appendix. We include phase retrieval examples in Figure 1 because we see it as the hardest inverse problem considered in DAPS.

However, as this is listed as a major weakness, we note that we do not present an atypical number of visual comparisons compared to the baselines. For example, the DPS paper also contains two such comparisons in the main body, plus a single figure to highlight the failure mode of a specific baseline in the supplemental material, and DAPS only has one such figure in the main body and two figures (12 and 28) for the CS-MRI task in the Appendix.

If you have a specific concern about our method, whether it be image quality or diversity on a specific task, and would like to see a concrete visual comparison across methods to address this specific concern, please let us know and we would be happy to add it.

(W6, published DAPS numbers) We report DAPS-1K for all tasks, while the DAPS paper reports DAPS-4K figures in Table 1 for nonlinear tasks. We attempted to say this in Appendix A.6.1:

(lines 1011-1014) ...the authors’ official public implementations and recommended task-specific hyperparameters are used throughout the benchmark. DAPS uses the DAPS-4K configuration for nonlinear tasks and DAPS-1K for linear tasks. For consistency in computational budget, PC-DAPS uses the DAPS-1K configuration for both linear and nonlinear tasks

but realize that the way this is written is not clear. We propose the following correction:

(lines 1011-1014) ...the authors’ official public implementations and recommended task-specific hyperparameters are used throughout the benchmark, with one exception: we report DAPS-1K results for all tasks. The original DAPS paper [10] states in Section 4.1 that linear tasks report DAPS-1K results and nonlinear tasks report DAPS-4K results. So that both (i) a consistent computational budget is used for a fixed algorithm across all tasks and (ii) DAPS and PC-DAPS are afforded comparable compute (as visualized in Figure 8), we report DAPS-1K results on all tasks.

We note that our reported DAPS SSIM and LPIPS closely match those in DAPS++ [2], which also runs DAPS-1K on all tasks (but reports FID instead of PSNR). Our numbers are also not uniformly lower than the DAPS paper’s numbers: they are better on 9 of 30 linear results and 4 of 18 nonlinear results, despite using a quarter of the NFEs on the latter set. We also found that DAPS metrics vary noticeably across runs and machines, as is acknowledged by the DAPS authors in their GitHub issues page.

AuthorsOpenReview ↗

Response to KXbq (3 of 3)

(W6, DPnP ImageNet) We observed this fact and have analyzed it in detail. For the phase retrieval task, a sample can be of high quality modulo Fourier ambiguity if it is one or more color channel inversions and/or a 180$^\circ$ rotation away from a high-quality sample. We attribute the combination of low PSNR yet high SSIM to samples with one or more inverted color channels but which have the correct orientation: SSIM of the right-side-up image is relatively high because the structure of the image is correct, but PSNR is low because pixel values are significantly incorrect.

We observe DPnP often produces upright but color-inverted outputs, thus when we take best-of-kk over a run of four there is typically some sample with relatively high SSIM, but these samples still have low PSNR due to pixel value error. We validate this metrics mismatch behavior on an ImageNet validation image in the table below.

Channels flippedPSNRSSIM
0∘0^\circ180∘180^\circ0∘0^\circ180∘180^\circ
None∞\infty14.261.0000.362
One (red)18.4713.940.7820.366

(Q1) Yes, pruning is what prevents flipped outputs. We evaluate this with an ablation on FFHQ phase retrieval (table below). When we turn pruning off (mixture only), the rate of flipped samples increases from 1.5% to 11.0% and the rate of faithful samples decreases from 97.5% to 84.3%. Relative to the reference sampler, the mixture alone reduces the rate of flipped samples from 16.0% to 11.0%, and relative to reference w/ pruning the mixture further reduces them from 3.0% to 1.5%.

For the prior score in Eq. (9), the score separates invalid from valid samples with an AUROC of 0.95. Relative to valid (upright) reconstructions in FFHQ, the mean prior score is 69% worse for color-inverted samples and 21% worse for rotated samples. Overall, we find that no valid reconstructions are pruned.

MethodFaithful (↑\uparrow)Upside-down/color-inverted (↓\downarrow)
DAPS-1K + ODE schedule + Langevin drop78.8%16.0%
+ Mixture84.3%11.0%
+ Pruning95.0%3.0%
+ Pruning + mixture (PC-DAPS)97.5%1.5%

(Q2) The goal is to encourage exploration in pixel-space directions where the images in the bag vary, which we see as an improvement over encouraging exploration in all pixels isotropically. For example, if I have a photo of someone against a solid blue background, there will be large uncertainty in the bag for facial details but low uncertainty in the bag for background details. The isotropic DAPS approximation encourages prior-driven exploration in all pixels equally, while our approximation would encourage exploration in the face pixels and discourage exploration in the background pixels.

(Q3) Our covariance is task-dependent, as different tasks will have more or less sample diversity depending on how informative the measurement is. In contrast, the DAPS covariance schedule uses the forward diffusion noise schedule σt2\sigma_t^2 as its covariance proxy for all tasks, and so does not capture this task-dependent diversity. We used a variance floor of 10−410^{-4} for our experiments.

(Q4) Yes, on average the rotated outputs get a worse prior score on ImageNet as well. Relative to upright reconstructions, the mean prior score is 17% worse for rotated samples, which is a smaller gap than in FFHQ (21%). We note that, because the upright and flipped images are equally plausible under the measurements, we can only expect the posterior sampler to treat a flipped image as equally plausible iff the off-the-shelf prior score does. Thus, since the pretrained prior scores do not treat flipped samples as equally plausible (as observed in the numbers above), we can’t expect the resulting posterior sampler to treat them as equally plausible (even if we as humans semantically find them equally plausible).

(Q5) To compare methods on the same protocol, we could do best-of-kk vs. best-of-kk or mean vs. mean. We defer the mean vs. mean comparison to the Appendix since no existing methods claim to perform well on average and may be designed/tuned with best-of-kk evaluation in mind, but are open to moving the table to the main body. We report mean PC-DAPS values (rather than best-of-kk) in Table 1 to show that our tractable sampling method, which doesn’t require oracle access, has better image quality on average than the best samples produced by existing methods.

(Q6) There is a single typo in our ImageNet results that we spotted after submission: the PSNR standard deviation of phase retrieval on ImageNet should show 6.1 instead of 1.1. We reproduced all other elements of the table as presented in our submission, including PSNR mean, SSIM mean$\pm$std, and LPIPS mean$\pm$std on the ImageNet phase retrieval task, and confirm this is the only typo.

Comments on the paper

3 comments
AuthorsOpenReview ↗

References

We appreciate all the feedback and questions from the reviewers. In an effort to keep reference numbers consistent across responses, we supply a single list of all references used in our comments.

[1] Gabriel Cardoso, Yazid Janati El Idrissi, Sylvain Le Corff, and Eric Moulines. Monte Carlo guided denoising diffusion models for Bayesian linear inverse problems. In International Conference on Learning Representations (ICLR), volume 2024, pages 44001–44037, 2024.

[2] Hao Chen, Renzheng Zhang, and Scott S Howard. DAPS++: Rethinking diffusion inverse problems with decoupled posterior annealing. In European Conference on Computer Vision (ECCV), pages 223–240. Springer, 2026.

[3] Zehao Dou and Yang Song. Diffusion posterior sampling for linear inverse problem solving: A filtering perspective. In International Conference on Learning Representations (ICLR), volume 2024, pages 12069–12095, 2024.

[4] J. R. Fienup. Phase retrieval algorithms: a comparison. Appl. Opt., 21(15):2758–2769, Aug 1982. doi: 10.1364/AO.21.002758.

[5] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 257–265, 2016.

[6] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International conference on learning representations (ICLR), 2023.

[7] Jan Pawel Stanczuk, Georgios Batzolis, Teo Deveney, and Carola-Bibiane Schönlieb. Diffusion models encode the intrinsic dimension of data manifolds. In Forty-first International Conference on Machine Learning, 2024.

[8] Phong Tran, A. Tran, Quynh Phung, and Minh Hoai. Explore image deblurring via encoded blur kernel space. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11951–11960, 2021.

[9] Luhuan Wu, Brian Trippe, Christian Naesseth, David Blei, and John Cunningham. Practical and asymptotically exact conditional sampling in diffusion models. In Advances in Neural Information Processing Systems (NeurIPS), volume 36, pages 31372–31403. Curran Associates, Inc., 2023. doi: 10.52202/075280-1363.

[10] Bingliang Zhang, Wenda Chu, Julius Berner, Chenlin Meng, Anima Anandkumar, and Yang Song. Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20895–20905, 2025.

[11] Zezhong Zhang, Feng Bao, and Guannan Zhang. Iensf: Iterative ensemble score filter for reducing error in posterior score estimation in nonlinear data assimilation. Journal of Computational Physics, page 115357, 2026.

AuthorsOpenReview ↗

Rebuttal follow up

Hello reviewers,

We thank you again for your feedback, comments, and recommendations. As there are only a few days left in the author-reviewer discussion period, we wanted to follow up to ask if we have addressed your concerns, or if you have any remaining questions or concerns that we can address.

AuthorsOpenReview ↗

Rebuttal version uploaded

Dear Reviewers,

We thank you for your additional feedback. We have uploaded a rebuttal version of the paper which (i) includes new appendices containing additional requested figures and ablations from the review period, (ii) adds main body references to the new appendices where appropriate, and (iii) moves the mean-of-kk comparison from appendix to the main body (Table 3), as requested by reviewer KXbq. All new content (as well as minor grammatical and ordering edits made for clarity) are emphasized in the rebuttal version as blue text.

We note that the main body uses the 10th page made available for the discussion period, and that minor writing edits were made to the abstract but which do not fundamentally change its content (e.g., to highlight the new diversity ablations and the emphasis on nonlinear inverse problems).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.