acceptodds
Under review as a conference paper at ICLR 2027

Error-Conditioned Generative Solvers for Ill-Posed Inverse Problems

13 views

Abstract

Supervised conditional diffusion models solve inverse problems in a single sampling run with the measurement as conditioning, but never check whether the reconstruction reproduces the measurement under the forward operator. Plug-and-play posterior samplers enforce the measurement by optimizing a data-fidelity objective, which requires differentiating the operator and often the denoiser at every step, needs per-problem tuning, and is an unreliable proxy for reconstruction error when the operator has a large null space. We introduce error-conditioned generative solvers: at every sampling step the denoiser proposes a clean field, the forward operator is applied to it, and the resulting residual, together with a physics residual when governing equations are known, is fed back to the same denoiser as an input. Trained with the standard denoising loss, the network learns how to act on its own residual, with no backpropagation through the operator, denoiser, or sampling chain. We evaluate on six severely underdetermined problems: a new benchmark of three compressible flows reconstructed from an initial state and radiographs, together with compressed-sensing MRI, sparse-view CT and inverse scattering. Against the same conditional model at an equal number of network evaluations, residual feedback reduces reconstruction error by –. Our method attains the lowest reconstruction error of any posterior sampler we evaluate, while running – faster than plug-and-play samplers. Its samples preserve shock structure that a lower-error regression model smears, and of its residual-driven correction lies in the operator's null space, where the measurement is blind.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
inverse problems, diffusion models, posterior sampling, conditional generative models, scientific machine learning, computational imaging, partial differential equations

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_u7x1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an error-conditioned diffusion model for solving ill-posed inverse problems. At each diffusion step, a denoiser first predicts a clean estimate, computes its residual, and then uses this residual as additional conditioning in a second evaluation of the same denoising network. Both evaluations are trained using a supervised denoising objective, avoiding differentiation through the forward operator during inference. The method is evaluated on six inverse problems, including compressible flows, MRI, CT, and inverse scattering. The authors report improvements over conditional diffusion and several plug-and-play sampler solvers in reconstruction accuracy and inference time.

Critical strengths
  1. The paper presents consistent empirical improvements across six inverse problems. Importantly, the comparison against conditional diffusion uses the same backbone neural network and an equal number of network evaluations, making the reported gains fair.

  2. Additionally, the zero-residual ablation provides useful evidence that improvements arise from residual conditioning rather than simply introducing a second denoising evaluation. The analysis of corrections in the operator's null space is also an interesting attempt to characterize the learned mechanism.

Critical weaknesses
  1. The central idea of feeding measurement or physics residuals into a learned reconstruction network is well established in recurrent inference machines, unrolled reconstruction methods, and error-conditioned neural solvers. The main methodological difference here is incorporating this mechanism into a two-pass diffusion denoiser with shared parameters. While this is a reasonable design, I find the conceptual advance relatively modest. The paper primarily demonstrates that an existing residual-conditioning principle can improve conditional diffusion, rather than establishing a substantially new approach to inverse inference.

  2. The method requires paired training data and operator-specific training, which limits practicability, as each inverse problem requires training a diffusion model. The experiments demonstrate some robustness within a predefined measurement family, but do not establish transfer to fundamentally different forward operators. This limits the generality of the contribution relative to reusable diffusion priors, as it requires retraining for each inverse problem, leading to high computational costs.

Questions for authors

How does the proposed method perform on unseen acquisition geometries or forward operators? Is retraining necessary?

3 comments

Conceptual advance relative to RIMs, unrolled networks and error-conditioned solvers. We agree that feeding a residual to a network is an established principle, and we cite it as such. RIMs, unrolled networks and ENS produce a deterministic point estimate, and our contribution is to use the residual inside a posterior sampler. Please see "Concurrent work and point-estimation methods" in the general response for more detail. §6.3 also shows that the method uses the residual differently from a learned data-consistency step. 85% of the residual-driven correction lies in the null space of AA. A data-consistency step cannot make such a correction, because its update A⊤(Ax−y)A^\top(Ax - y) always lies in the row space of AA.

Operator-specific training and unseen operators. Like conditional diffusion, our method is designed for problems where the same family of measurements recurs and the requirement for accuracy justifies a one-time training run, rather than for transfer across operator families. A new operator family therefore needs a new model, while new view counts and dose levels within a family do not (App. A.8). Please see "Problem setting: one model per measurement family" in the general response for more detail and for the settings this covers.

Reviewer_u7x1OpenReview ↗

Response to authors

Thank you for your answers and clarifications. Regarding generalization, I appreciate the experiments in Appendix A.8 showing that a single model can handle different view counts and noise levels. However, I would be interested in understanding whether the proposed method can transfer to unseen acquisition geometries within the same operator family. For instance, could a model trained on one set of projection angles be evaluated on a different angular configuration without retraining? Since the method explicitly incorporates the forward operator and its residual during inference, I think this would be an interesting experiment to better understand its generalization capabilities. Additionally, I think the experimental evaluation could be strengthened by including more plug-and-play diffusion baselines, such as DiffPIR and DDNM, among other established solvers. This comparison would help better assess the advantages of the proposed method relative to existing approaches. Thank you again for addressing my concerns.

Thank you for the additional feedback.

Transfer to unseen angular configurations. Without retraining, our model transfers well to intermediate view counts it was not trained on, but not to angle configurations of a different kind, such as shifted, irregular or limited-arc angles (new App. A.8.5). Because the conditioning is computed in image space from the filtered and unfiltered backprojections, a trained model accepts any set of projection angles without changes to the architecture.

Nonetheless, with residual feedback, the model is more robust to these shifts than conditional diffusion. First, the model of App. A.8 is trained on 4, 16, 64, 256 and 1024 evenly spaced views. At the held-out counts in between, it has the lowest NRMSE at 32, 128 and 512 views and is as accurate as DPS at 8 views. In the tables below, bold indicates the best NRMSE and italics the second best in each column.

Method8 views32 views128 views512 views
DPS0.2320.1850.1540.132
Conditional diffusion0.2810.1870.1480.128
Ours0.2320.1730.1410.117

Second, we evaluate the same model on unseen angle configurations, a 4-view set shifted by 10°, 8 irregularly spaced angles and 16 views over a 90° arc. On these configurations, DPS is more accurate than both amortized models (ours and conditional diffusion). Relative to evenly spaced angles at the same view count, the NRMSE of ours increases 26–35% less than that of conditional diffusion. On the irregular 8-view set, for example, the NRMSE of conditional diffusion rises from 0.281 at 8 evenly spaced views to 0.535, and ours from 0.232 to 0.397. This suggests that recomputing the residual with the test-time operator compensates for part of the shift. Third, after we fine-tune the same model for 200,000 steps on random angle sets, it has a lower NRMSE than DPS on all three configurations. The paired 95% confidence interval excludes zero on the 90° arc and includes zero on the other two. It also remains more accurate than conditional diffusion fine-tuned the same way.

MethodShifted, 4 viewsIrregular, 8 views90° arc, 16 views
DPS0.2860.2730.239
Conditional diffusion, fixed angles0.4110.5350.599
Ours, fixed angles0.3590.3970.458
Conditional diffusion, random angles0.3250.3140.269
Ours, random angles0.2820.2690.228

These results are consistent with the scope of our method. As §7 states, the family of forward operators must be fixed before training, and App. F now states which angle sets a model trained on evenly spaced angles transfers to. Our general response discusses this scope in more detail ("Problem setting"). Within the family of geometries seen in training, our method is at least as accurate as DPS and about 10× faster. For geometries outside that family, a plug-and-play prior or training on a wider family is the better choice.

DiffPIR and DDNM. Following your suggestion, we added both to Table 3 on the three imaging tasks. They use the same unconditional prior as DPS, DAPS and PnP-DM and were tuned with the same protocol (App. D.2). Our method has a lower NRMSE than both on every task, and the paired 95% confidence intervals exclude zero (App. A.9). DDNM runs at about our speed, and DiffPIR about 5× slower. Both methods require the proximal map or pseudo-inverse of a linear operator. On Kelvin–Helmholtz and Rayleigh–Taylor the radiographs measure the partial density of the heavy fluid, ρ(1−Y), a product of two state channels, so the operator is not linear. On compressible Euler the operator is linear, but the projection step of both methods cannot include the physics term that the other plug-and-play samplers use on the flows (App. D.2).

MethodCS-MRICTScattering
DPS0.2810.2920.325
DiffPIR0.2930.3040.348
DDNM0.3470.2980.394
Ours0.2570.2340.333
Reviewer_mDxiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes an error-conditioned generative solver for severely underdetermined inverse problems. At each diffusion sampling step, the model first predicts a clean estimate, computes the measurement residual (and optionally a PDE residual), and performs a second denoising pass conditioned on these residuals. The method learns residual-based corrections without backpropagating through the forward operator or the sampling trajectory. Experiments on six inverse problems demonstrate improved reconstruction accuracy over conditional diffusion and several diffusion posterior samplers.

Critical strengths
  1. Simple and effective residual-feedback mechanism. The proposed two-pass design incorporates measurement consistency information without requiring test-time gradient-based optimization. The experiments demonstrate consistent improvements over conditional diffusion at matched network evaluation budgets, and the ablations support the effectiveness of residual feedback.
  2. Diverse experimental evaluation. The paper evaluates six inverse problems, including a challenging new fluid-flow reconstruction benchmark. The additional analyses of shock structures, sample diversity, and null-space corrections provide useful insights beyond standard reconstruction metrics.
Critical weaknesses
  1. Limited methodological novelty relative to closely related work. The core idea of using residuals to learn iterative corrections has been explored in RIMs and ENS. More importantly, ReCo-Diff employs a closely related two-pass residual-conditioned scheme, while Payeur et al. (2026) integrate recurrent inference into diffusion-based posterior sampling. Although these works are discussed, the paper does not provide direct experimental comparisons or sufficiently establish the advantages of its specific design over these approaches.
  2. Insufficient validation across acquisition regimes and of posterior sampling quality. The main medical imaging experiments focus on extremely sparse measurements (16× single-coil MRI and 4-view CT), limiting the practical generalizability of the conclusions. Although Appendix A.8 explores additional CT view counts, conventional reconstruction baselines are missing. Furthermore, the paper claims posterior sampling capability but provides no quantitative evaluation of posterior calibration.
Questions for authors
  1. Qualitative results. The reconstruction images in Figure A5 are too small to clearly distinguish the methods or appreciate the proposed method's advantages. Why is only the real component shown for CS-MRI instead of the magnitude image? Larger visualizations, zoomed-in regions, and magnitude reconstructions would make the comparison more informative.
  2. Undersampling settings. Why is CS-MRI evaluated only at 16× acceleration? While extreme undersampling is relevant to the paper's motivation, evaluating more commonly used acceleration factors (e.g., 4× and 8×) would help establish the method's applicability. Similarly, more comprehensive comparisons at moderate CT view counts would be valuable.
  3. CT results under dense sampling. In Table A6, even with 1024 views, the proposed method achieves only 27.82 dB PSNR and 0.801 SSIM. What accounts for the remaining reconstruction error under such dense angular sampling? Please consider reporting FBP or iterative reconstruction baselines and qualitative results at higher view counts. Clarification of the attenuation normalization and noise simulation would also be helpful.
  4. Comparison with related methods. Could the authors clarify the methodological differences from ReCo-Diff and diffusion-based RIM approaches, and provide direct comparisons or targeted ablations to establish the benefits of the proposed formulation?
  5. Posterior sampling validity. Beyond demonstrating sample diversity, is there quantitative evidence that the generated samples accurately represent the conditional posterior, such as uncertainty calibration or coverage analysis?
  6. PDE residual under unknown forcing. The flow datasets include unobserved stochastic body forcing, but the PDE residual appears to omit this forcing term. How much does this model mismatch affect the residual evaluation and the comparison with physics-informed baselines such as PINO and PINN?
5 comments

Methodological novelty and comparison with related methods. ReCo-Diff and Payeur et al. are concurrent work, which the submission already cited and discussed, and the ICLR 2027 reviewer guidelines state that the lack of comparisons to such work cannot be a basis for rejection. RIMs, unrolled networks and ENS produce a deterministic point estimate, so they address a different problem from posterior sampling. We nevertheless compared against ReCo-Diff on CT and MRI (App. A.4). Please see "Concurrent work and point-estimation methods" in the general response for more detail and the results. For a targeted ablation, the zero-residual control (App. A.5) separates the residual from the second pass. On compressible Euler, a second pass without the residual recovers about a third of the gain over conditional diffusion (0.1741 to 0.1620, against 0.1400 for ours).

Acceleration regimes, undersampling settings and conventional baselines. Following your suggestion, we added CS-MRI at 4×, 32× and 64× with separately trained models (App. A.11), and ours has the lowest NRMSE at every acceleration (0.153 at 4×, next best 0.170). We also added FBP and TV at every CT view count from 4 to 1024 (App. A.8). Ours has the lowest NRMSE at every view count, not only the sparsest, and TV is the closest conventional method (0.133 vs. 0.106 at 1024 views).

Posterior calibration and sampling validity. We added a calibration analysis of shock location on compressible Euler (App. A.3). The 90.5% and 52.4% central intervals of our 20 samples contain the true front on 88.3% and 52.8% of shock lines, with the lowest continuous ranked probability score (CRPS) of the samplers. The intervals of conditional diffusion are too wide (96.7%, 71.0%), and those of DPS, DAPS and PnP-DM too narrow (73–80% at 90.5%).

Qualitative results. We replaced the submitted Fig. A5 with one page-width figure per imaging task (now Figs. A8–A10 in App. A.12), with zoom insets, the MRI magnitude and shared-range color bars.

CT results under dense sampling. We added an account of this to App. A.8, with FBP and TV baselines and reconstructions at 256 and 1024 views. At 1024 views, accuracy is limited by the measurement noise rather than by the number of views. Without noise, unregularized least squares reaches 40.6 dB. On the noisy sinogram the same iteration diverges, since the dose is fixed per detector bin and view and the noise on each ray grows with its attenuation. On the same data TV and FBP reach 25.46 and 21.57 dB (ours 27.81 dB). We also added the attenuation units (cm$^{-1}$, with the 1282128^2 grid spanning 26 cm) and the noise model to App. C.

PDE residual under unknown forcing. We added the reason for the body force to App. C. The body force stands in for discrepancies between the simulation model and reality, so it is withheld from every method by design. Every physics-using method reads the same residual without the body force, so the comparison with PINO and the PINN is on equal terms. For our method the physics residual contributes little. Removing the physics residual from the input costs 0.6% NRMSE on compressible Euler (App. A.5).

Thank you again for your review. As the discussion period is ending soon, we wanted to check whether our response and the revised paper address your concerns. We added new experiments in response to your comments and would welcome any further questions or feedback.

Reviewer_mDxiOpenReview ↗

reply to the rebuttal

The response has addressed most of my concerns, and I am satisfied with the revisions. Given the response and the manuscript's current quality, I have decided to raise my score to 3.

I do not oppose accepting this paper, as I agree that it makes a meaningful methodological contribution. However, I have to say that the paper is still some distance away from being a high-quality work, mainly because of its experiments and presentation.

The paper is quite difficult to follow. Only after carefully reading the authors’ response and the newly added material did I fully understand some of the things the authors were trying to present. This also raised several new concerns. I hope the authors will consider them. Even if the paper is accepted, I think these issues should at least be clarified in the manuscript to avoid misleading readers.

  1. Experimental settings

I am not very familiar with PDE problems, but I have more experience with medical imaging. My concerns are mainly about two aspects.

First, the reconstruction resolution. It seems that both CT and MRI experiments are performed at only 128×128 resolution. To me, these are essentially toy experiments. I seriously doubt how much we can learn from such low-resolution results. They may show that the proposed method works better under extreme undersampling and has some useful properties. However, the results could be quite different when scaling up to normal image sizes, at least 256×256, and preferably 320×320 for MRI and 512×512 for CT. Why not conduct experiments at normal resolutions in the first place? Many diffusion-based inverse problem solvers published at ICLR, ICML, and NeurIPS already do this. It is not an especially difficult requirement. From the current CT and MRI qualitative results, I can hardly see any useful details.

Second, the undersampling settings. The main experiments focus on extremely undersampled cases, such as 4-view CT and 16× MRI acceleration. The main qualitative figures also show results under these extreme settings. Although the authors have added experiments with other sampling rates in the appendix, the main results still focus heavily on these extreme cases.

Combined with the 128×128 reconstruction resolution, these settings have very little practical meaning for medical imaging. In real scenarios, we do not need such extreme undersampling at the cost of making the images nearly unusable. I would rather see more attention to how well the proposed method works under reasonable acquisition settings with practical value. Simply comparing PSNR numbers among a set of poor reconstructions does not tell us much. Otherwise, I think standard tasks such as natural image super-resolution would be more suitable for evaluating the algorithm itself.

  1. CT noise simulation

I have a further concern about the CT noise simulation in Appendix A.8 and Appendix C.

The authors use a post-log Gaussian approximation with variance (\exp(p)/I_0), rather than directly sampling Poisson photon counts as in LoDoPaB-CT. When the incident photon count is large, this approximation should be reasonably accurate. However, at (I_0=4096), it may be inaccurate for rays with very few transmitted photons.

I am also concerned about directly adopting (I_0=4096) from LoDoPaB-CT. The key point is that the CT acquisition geometries discussed in the two papers are quite different. LoDoPaB uses 513 detector bins, while this paper uses only 128. The same incident photon count per detector bin does not correspond to the same physical exposure or dose under different acquisition geometries. The low-dose simulation in this paper may therefore be too extreme.

  1. Overall concern

Overall, especially considering the reconstruction resolution, undersampling settings, and CT noise simulation, I think the medical imaging experiments have important limitations.

These experiments can still demonstrate that the proposed method makes a contribution to severely underdetermined inverse problems. In other words, the proposed method may indeed be a strong approach for different types of inverse problems. I do not question its methodological value.

However, the experimental settings are quite far from realistic medical imaging conditions. This raises concerns about the method's practical applicability and may also give readers a misleading impression.

This is particularly important for a paper published at ICLR. Readers unfamiliar with medical imaging may follow these experimental settings in their future work without realizing how far they are from realistic acquisition conditions. I do not think this would be good for the development of the research community.

I have raised my score because the authors have addressed my previous concerns and the paper has a meaningful methodological contribution. Still, I strongly encourage the authors to improve the experimental presentation and clearly state the limitations of these medical imaging settings in the final paper.

Reviewer_19p8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors train a conditional diffusion model that, similar to self-conditioning, receives its own output and an additional error operator as input. This conditional model is trained in a supervised manner for each task, using a paired dataset of ground truth images and observations.

The authors introduce new flow-field benchmark tasks (compressible Euler flow, Kelvin–Helmholtz, and Rayleigh–Taylor instabilities). They further evaluate their method on sparse-view CT (4 angles) and accelerated MRI.

Critical strengths
  • The proposed method is a simple and straightforward modification of standard conditional diffusion training (only two additional lines in Algorithm 1), and it appears to improve performance across a range of tasks.
  • The authors introduce new benchmark tasks, which could be a useful contribution to the community.
Critical weaknesses
  • In the introduction, you write that the ''posterior is broad and often multimodal, so its mean is typically not a plausible field'' (line 37). However, the main results only report pointwise metrics. The sample spread shown for the shock profiles is a nice addition, and it would be great to also see multiple samples for the CT and MRI tasks.
  • The method is supervised, which means it has to be retrained for every new setting (e.g., every new acquisition geometry in CT) and requires paired data for each of these settings. In contrast, many of the current diffusion-priors only need ground truth images and can be adapted to many different forward operators.
  • I found Section 6.3 difficult to follow on its own. For example, Δattr\Delta_{attr} it not clearly defined. It is also not clear where the statement "DPS is 70% row-space" comes from. Is this the mean over all sampling steps? Moving some of the key definitions into the main text would make this section more accessible.
  • A resolution of 128x128 is quite low for CT and MRI and not very representative of realistic settings. In Figure A5, for example, many details already appear pixelated in the ground truth. For comparison, the recent DM4CT benchmark [1] uses 512x512 images.
  • The results (e.g., Tables 2 and 3) do not include standard deviations.

Minor points:

  • Several parts of the appendix are never referenced in the main text, e.g., Appendices A.4, A.5, and A.8.
  • The authors sometimes refer to the nullspace as N(A)\mathcal{N}(A) and sometimes null(A)\text{null}(A)
  • In Line 126 it should be A(x+v)=A(x)A(x + v) = A(x) for v∈N(A)v \in \mathcal{N}(A).

[1] Shi et al. “DM4CT: Benchmarking Diffusion Models for Computed Tomography Reconstruction.” ICLR 2026.

Questions for authors
  1. The benchmark details in Appendix C do not specify the type and level of observation noise. For instance, what kind of noise was used for the flow examples or for MRI?
  2. Line 139: Why does the PDE residual not vanish for the ground-truth data?
  3. Could you add color bars to the error and residual maps? Do all error maps use the same color range?
  4. Section 6, Baselines: Do all conditional models have the same training budget?
  5. Why is the posterior mean in Table 2 computed for only 5 test items? Would it be possible to compute it on the full test set? In addition, 20 samples seems rather few for estimating the posterior mean.
  6. In Appendix A.8 (“Robustness to Measurement Degradation”), is the model retrained for each experimental setting (noise level, number of angles), or is training amortized across all settings?
4 comments

Multiple samples for CT and MRI. Following your suggestion, App. A.12 now has one figure per task with four samples of ours and their per-pixel standard deviation, and a figure comparing samples of ours, conditional diffusion and DPS on the same item.

Supervision and retraining. This concerns the problem setting the method targets. Please see "Problem setting: one model per measurement family" in the general response.

§6.3 and "DPS is 70% row-space". We revised §6.3 to clearly define terms. The revised section defines the visible fraction frow(v)=∥Pv∥2/∥v∥2f_{\mathrm{row}}(v) = \|Pv\|^2/\|v\|^2 with P=A+AP = A^+A, the shuffle control Δshuf\Delta_{\mathrm{shuf}} (the correction recomputed at the same state with another sample's residual), and Δattr=Δ−Δshuf\Delta_{\mathrm{attr}} = \Delta - \Delta_{\mathrm{shuf}}. It also specifies how the 70% figure for DPS is computed. The value 0.70 is frowf_{\mathrm{row}} of the DPS guidance direction, energy-weighted over all sampling steps at the DPS step size before retuning.

128×128 resolution. We chose this resolution so that every method could be trained and tuned on all six tasks with reasonable compute.

Standard deviations. We added 95% bootstrap confidence intervals for every learned method in Tables 2 and 3, and paired intervals against each baseline (App. A.9). Please see the general response for a summary.

Minor points. Thank you for catching these. App. A.4, A.5 and A.8 are now referenced from the main text, null(A) is used throughout, and line 126 of the submission (line 132 of the revision) now reads A(x+v)=A(x)A(x+v) = A(x) for v∈null(A)v \in \mathrm{null}(A).

Type and level of observation noise. We added the noise models to App. C. On the flows, the observed first frame carries a fractal density perturbation of amplitude 0.3, and the radiographs are noise-free (App. A.8.4 adds radiographic noise). CS-MRI adds complex Gaussian noise to the sampled k-space at 30 dB SNR, CT uses the Gaussian approximation of log-Poisson counts at I0=4096I_0 = 4096, and inverse scattering adds Gaussian noise with σ=10−4\sigma = 10^{-4}.

Why the PDE residual does not vanish on the ground truth. We expanded App. B.3 to explain this more clearly. The residual is a single finite-difference pass, fast enough to evaluate at every sampling step but approximate. The stored frames are far apart for a one-frame time difference (CFL per frame 3.3, 24 and 31), the differential form does not hold at shocks, and the unobserved body force is not included. Driving such a residual to zero would pull a reconstruction away from the true field. Our method reads the residual as an input instead and learns from data what a nonzero value means, which lets our method use a cheap, approximate residual effectively.

Color bars and color ranges. We added color bars to every figure with error or residual maps. All methods in a row share one color range.

Training budget. Yes. Every model trains for the same number of steps at the same batch size and learning rate, which App. D.1 now states.

Posterior mean on 5 items, and 20 samples. The posterior mean is expensive to evaluate across all methods. Even 5 items with 20 samples each takes 100 sampling runs per method, so we report it on 5 items as a reference for what averaging gains, not as one of the compared methods. Twenty samples appears to be sufficient for the mean. Going from 10 to 20 samples changes the NRMSE of the mean by at most 0.003 for ours and 0.009 for any posterior sampler, which we added as App. A.10. For CT and MRI, App. A.4 also reports the mean of 20 samples on 25 items.

Retrained or amortized in App. A.8. Amortized. One checkpoint of ours and one of conditional diffusion cover all five view counts, and another pair the three dose levels. The view count is not given to the model.

Thank you again for your review. As the discussion period is ending soon, we wanted to check whether our response and the revised paper address your concerns. We added new experiments in response to your comments and would welcome any further questions or feedback.

I thank the authors for their detailed response and the substantial revisions. Most of my concerns have been addressed. A single checkpoint covering 4 to 1024 views without being told the setting is a useful result, and it partly answers my concern about retraining for every acquisition geometry.

Resolution. The response justifies the resolution 128x128 by compute, but this does not really resolve my concern. At this resolution most of the interesting and important anatomical details are lost, see e.g. Figure A13 in which the images are very pixelated. Similar for the MRI results in Figure A12.

Also, the title of Table A7 Variable-view sparse-view CT at fixed dose is misleading. Given that I0I_0 is fixed, the total dose increased with the number of measured angles. Correct would be that the dose of each ray is fixed.

Given the revisions and the methodological contribution, I’m raising my score to 3.

Thank you for the additional feedback.

Medical imaging resolution. We agree that much of the clinically relevant anatomical detail is lost at 128×128. App. F now describes the CS-MRI and CT tasks as benchmarks of severely underdetermined reconstruction rather than clinical protocols, notes that 128×128 is below clinical resolution, and states that the results are not evidence of clinical performance. The limitations in §7 refer to App. F.

Title of Table A7. Thank you for catching this. Since I0I_0 is fixed per detector bin and view, the total dose grows with the number of views. The title now reads "fixed dose per ray", and the caption and App. A.8.1 state that the total dose grows with the number of views.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response

We thank the reviewers for their feedback. The revision adds the analyses and baselines the reviewers asked for. Changes in the PDF are highlighted in blue.

Major changes

  • Confidence intervals for every result in Tables 2 and 3 (App. A.9). 95% bootstrap intervals over the test items for every learned method, and paired intervals of our NRMSE minus each baseline's. Ours is lower than conditional diffusion, DAPS and PnP-DM on every task, lower than DPS on four, and within the interval of DPS on Kelvin–Helmholtz and inverse scattering.
  • Posterior calibration on shock locations (App. A.3). On compressible Euler, the 90.5% central interval of our 20 samples contains the true front on 88% of shock lines, the closest to nominal of the samplers.
  • CS-MRI at 4×, 32× and 64× (App. A.11) and FBP and TV at every CT view count from 4 to 1024 (App. A.8). Ours has the lowest NRMSE in every setting.
  • Comparison with ReCo-Diff (App. A.4), adapted to our backbone, data and operators on sparse-view CT and CS-MRI (below).
  • Showing multiple samples for CT and MRI (App. A.12), and the submitted Fig. A5 split into one larger figure per imaging task (now Figs. A8–A10).
  • Plug-and-play baselines retuned per task (App. D.2), which lowered DPS's error on the flows and inverse scattering. Conditional diffusion with DPS guidance is in App. A.2.
  • Clarifications of the observation noise (App. C), the training budget (App. D.1) and the definitions in §6.3, and a list of application settings (App. F).

Problem setting: one model per measurement family (19p8, u7x1). Like conditional diffusion, our method is trained for one family of measurements. It is designed for problems where the same kind of measurement recurs and the requirement for accuracy justifies a one-time training run, such as clinical imaging protocols, radiography at an experimental facility, or industrial inspection (App. F). We do not claim transfer across operator families. When the operator changes arbitrarily from case to case, a plug-and-play prior is the better fit. Within a family, our method needs only one model. A single checkpoint handles every CT view count from 4 to 1024, and another every dose level, without being told the setting (App. A.8). In return for per-task training, inference is 9–13× faster than the plug-and-play samplers, with lower NRMSE on four of six tasks and comparable NRMSE on the other two.

Concurrent work and point-estimation methods (mDxi, u7x1). RIMs, unrolled networks and ENS produce a deterministic point estimate, whereas our method samples the posterior, so they address a different problem. ReCo-Diff (arXiv, March 2026) and Payeur et al. (arXiv, July 2026) are concurrent work. The submission already cited and discussed both (§3), and the ICLR 2027 reviewer guidelines state that the lack of comparisons to such work cannot be a basis for rejection. We nevertheless adapted ReCo-Diff to our setting and compared against it (App. A.4, D.2). Like regression, ReCo-Diff is a deterministic estimator trained with squared error, so we compare it with regression and with the mean of our samples. ReCo-Diff's single output is more accurate than one sample of ours (NRMSE 0.221 vs. 0.241 on CT, 0.251 vs. 0.261 on MRI). The mean of 20 of our samples is more accurate on CT (0.195) and within the confidence interval on MRI (0.244), and ReCo-Diff's outputs have less than half the high-frequency energy of our samples.

AuthorsOpenReview ↗

Additional revisions changelog

  • Kelvin–Helmholtz and Rayleigh–Taylor: corrected the physics residual input of our method at evaluation, and on Rayleigh–Taylor added gravity to the residual, matching the physics terms of PINO+TTOP and DPS. Our NRMSE is now 0.116 and 0.127 (submitted 0.129 and 0.151). App. A.8.4 keeps the earlier Rayleigh–Taylor model and says so.
  • PINO+TTOP on Kelvin–Helmholtz and Rayleigh–Taylor: corrected its radiograph term, which lowered its NRMSE (0.510 → 0.154, 0.350 → 0.210).
  • Sparse-view CT, CS-MRI and inverse scattering: every method is now evaluated on the same fixed noise realization per item (NRMSE changes of at most 0.0035).
  • The zero-residual control now uses the same checkpoint selection as every other model (NRMSE 0.1674 → 0.1620).
  • App. A.4: the regression figure and its numbers now consistently refer to Kelvin–Helmholtz.
  • App. A.9 states that the PINN's entries on Kelvin–Helmholtz and Rayleigh–Taylor are over the items where it converged (47 and 39).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.