acceptodds
Under review as a conference paper at ICLR 2027

MPR-Flow: Efficient Inverse Problem Solving for Sequential Measurements

Abstract

Many real-world imaging techniques rely on combining multiple measurements acquired sequentially. The final image can then be obtained by solving an inverse problem that accounts for the physical model of the acquisition process. In many applications, it is essential to minimize the number of data acquisitions to avoid multiplying time-consuming, costly, or risky procedures. Being able to obtain the best possible reconstruction as measurements are acquired is thus critical to avoid unnecessary acquisitions. In this setting, each reconstruction can be a useful computation state for the next step. With this in mind, we introduce Measurement-Progressive Reconstruction Flow (MPR-Flow), a plug-and-play flow solver that exploits continuity at two scales. First, across acquisition stages, previous reconstructions are used as a prior to speed up the reconstruction. Second, across solver iterations, the re-noising step is constrained so that well-measured regions are pre- served. MPR-Flow needs no retraining, nor any changes in the prior. Experiments on simulated and real datasets show that it outperforms state-of-the-art flow-based inverse solvers in terms of reconstruction quality and acquisition time.

Primary area:
generative models
Keywords:
Plug-and-Play, Flow Matching, image restoration, inverse problems, sequential sensing

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_8twwOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MPR-Flow reconstructs a fixed image as measurements accumulate. It combines acquisition-dependent warm starts with reflected noise coupling, which refreshes noise in measured directions and reverses it in the measurement nullspace. Experiments on progressive inpainting, random sensing, synthetic DNA AFM, and retrospective MRI report substantial improvements over independently restarted flow solvers at matched per-stage network budgets, alongside runtime savings.

Critical strengths
  • Substantial empirical gains across several operators without retraining the prior. With ten evaluations per stage, MPR-Flow improves 30% inpainting from 26.1 to 32.0 dB and noiseless MRI from the strongest baseline’s 32.2 to 36.1 dB. Shared priors, component ablations, and measured runtime comparisons make the efficiency results promising.
Critical weaknesses
  • The contribution beyond straightforward continuation is insufficiently isolated. All main baselines restart at every acquisition. It shows that ADWS accounts for most of the improvement, e.g., for random sensing, 14.7 to 24.2 dB, versus 25.1 dB for the complete method. There is no tuned fixed-start continuation control or empirical comparison with continued FLOWER, despite Appendix A.2 describing that adaptation. Consequently, the benefit of the proposed acquisition-dependent schedule over simpler state reuse remains unclear.

  • The data-consistency update is not reproducible as written. The fidelity divides by measurement-noise variance, yet experiments include zero noise, where this expression is undefined. For noisy experiments, the reported step schedule also lacks an explicit normalization compensating for this variance scaling. This may be an omitted implementation convention rather than an implementation error, but it needs clarification because it directly determines the algorithm.

Questions for authors
  • Clarification: What exact fidelity loss, noise floor, and gradient scaling were implemented, particularly for MRI with sigma=0 and 0.03?

  • Assessment-critical comparison: How does MPR-Flow perform against tuned constant-time continuation and ADWS applied to FLOWER under matched computation? These comparisons would materially affect my assessment of the methodological contribution.

4 comments
AuthorsOpenReview ↗

W1 + Q2. Isolation of contribution beyond continuation.

We thank the reviewer for allowing us to clarify this point. First, we agree that warm start brings the biggest performance increase. However, as detailed in the theoretical guarantees added to the appendix, warm start shows great synergy with RNC, which is illustrated by the consistent gains brought by ADWS+RNC compared with ADWS alone cf. Tab. 2 of the manuscript: +1 PSNR on inpainting 10%, +1.8 on inpainting 30%, 0.8 PSNR on noisy MRI. To further illustrate ADWS's good properties, we also added the following comparison between fixed warm start and ADWS both including RNC and using the settings from Tab. 2.

MetricInpainting10%Inpainting30%Random matrixMRI0MRI0.03
Fixed WS (PSNR / SSIM)23.816 / 0.740531.762 / 0.936324.537 / 0.716335.961 / 0.876935.413 / 0.8925
ADWS (PSNR / SSIM)24.589 / 0.778531.962 / 0.940925.136 / 0.734736.075 / 0.878035.495 / 0.8933
PSNR gain (mean ± SD)+0.7729 ± 0.6695+0.2001 ± 0.1958+0.5999 ± 0.3176+0.1136 ± 0.2325+0.0819 ± 0.1484
SSIM gain (mean ± SD)+0.03795 ± 0.02135+0.00466 ± 0.00290+0.01840 ± 0.01651+0.00113 ± 0.00249+0.00082 ± 0.00179

To further detail the gain brought by ADWS, we present the following ablation comparing PnP-Flow with a fixed warm start, with and without RNC, to MPR-Flow.

MethodInpainting 10%Inpainting 30%Random matrixMRI σ=0MRI σ=0.03
PnP-Flow16.290 / 0.391726.113 / 0.794214.653 / 0.303030.563 / 0.654230.591 / 0.6714
PnP-Flow + fixed start22.952 / 0.694230.658 / 0.917924.025 / 0.696235.270 / 0.890934.577 / 0.8864
PnP-Flow + fixed start + RNC23.816 / 0.740531.762 / 0.936324.537 / 0.716335.961 / 0.876935.413 / 0.8925
MPR-Flow24.589 / 0.778531.962 / 0.940925.136 / 0.734736.075 / 0.878035.495 / 0.8933

ADWS shows consistent gains over a fixed warm start. ADWS brings more gains on inpainting 10% (+0.77 PSNR) and random matrix unmixing (+0.6 PSNR). MPR-Flow improves PSNR over fixed time warm start by 1.64, 1.30, 1.11, 0.81, and 0.92 dB for each of the inverse problems respectively.

In addition we expanded our experiments to FLOWER under matched computation.

MethodInpainting10%Inpainting30%Random matrixMRI0MRI0.03
FLOWER15.487 / 0.359325.011 / 0.737813.753 / 0.254732.239 / 0.844332.094 / 0.8495
FLOWER + warm start21.687 / 0.634230.160 / 0.908421.076 / 0.599535.114 / 0.886634.533 / 0.8870
FLOWER + ADWS22.376 / 0.679330.407 / 0.916121.127 / 0.601135.275 / 0.887834.689 / 0.8885
FLOWER + ADWS + RNC23.090 / 0.714431.306 / 0.929621.643 / 0.634535.804 / 0.869835.360 / 0.8894
MPR-Flow24.589 / 0.778531.962 / 0.940925.136 / 0.734736.075 / 0.878035.495 / 0.8933

Here also, the same conclusion remains: warm start brings large gains compared to Flower only, while ADWS and RNC both add consistent gains over each task.

AuthorsOpenReview ↗

W2 + Q1. Concerns about the data consistency step

Thank you for identifying this inconsistency. In the manuscript the data-fidelity term was incorrectly reusing the noise-normalized fidelity from the inverse-problem introduction. The implementation, uses the the unnormalized least-squares fidelity to compute the data fidelity.

F~s(x)=12∥Asx−ys∥22,\widetilde F_s(x)=\frac12\|A_sx-y_s\|_2^2,

with gradient

∇F~s(x)=As∗(Asx−ys).\nabla\widetilde F_s(x)=A_s^\ast(A_sx-y_s).

The implemented data-consistency update is therefore a gradient step in the direction of the data consistency

x_{s,n} - \eta(\tau_{s,n}) A_s^\ast(A_sx_{s,n}-y_s).$$ There is no inverse-variance factor, variance floor, or additional gradient normalization applied, and the same expression is used for noiseless and noisy measurements, including both MRI problems.
Reviewer_8twwOpenReview ↗

Response to the rebuttel

Dear authors,

Thank you for the clarification and additional experiments.

The unnormalized data-fidelity formulation resolves my concern afbout the update specification.

The added fixed warm-start and continued FLOWER comparisons support incremental benefits from ADWS and RNC, although the acquisition-dependent schedule contributes relatively modest gains in MRI. One clarification remains useful: whether the fixed starting time was independently tuned using a comparable search budget. Could you clarify how the fixed starting time was selected and whether it received comparable tuning.

AuthorsOpenReview ↗

Clarification on the warm start selection

Dear Reviewer,

Thank you for your response, the fixed starting time was tuned and selected using the same grid as ADWS, as we now clarify on Table 3, for both alpha and time with the same parameter grid was used on a different held out set of 128 images.

In general we observe the best performing value of the fixed warm start solver be equal or lower than the best performing value for ADWS, with the biggest differences in PSNR coming from the problems where the values of the warm starting time differ, which might hint at some of the benefit being due to ADWS allowing later stating times to retain stability.

Also we suspect the reason why we might not see a big difference for the MRI problem might be due to the fact that the initial MRI measurement carrying more information, due to the initial calibration contributing 31 lines in the k space, while subsequent stages add only 10 each.

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

The paper presents a new flow matching-based reconstruction algorithm that is designed for scenarios where measurements are observed sequentially, and the goal is to provide intermediate reconstructions for every new incoming measurement. The paper proposes an extension of the PnP-Flow algorithm to take into account this sequential setting, which i) modifies the starting time of the flow as more measurements are available, ii) designs a custom renoising process that uses iid noise in the observed parts (range space) and negatively correlated noise on the non-observed parts (nullspace). In a series of experiments on compressed sensing, inpainting and single-coil MRI problems, the proposed method is shown to outperform other flow-based baselines with cold-start.

Critical strengths
  • The paper tackles an interesting sequential sensing setting that is not often addressed by inverse problems papers
  • The idea of introducing negatively correlated noise across stages is interesting and novel (to the best of my knowledge).
  • The method seems to significantly outperform the cold-start flow baselines
Critical weaknesses
  • It is unclear why the proposed RNC approach works. I couldn't find any explanation that motivates this choice. As the paper's main contribution, I would expect at least some theoretical or practical intuition.

  • The paper has multiple typos and presentation errors, which make it harder to follow:

    • Something seems to be missing after Eq. 12 (after the comma)
    • \tau is missing subindices in Eq. 9
    • \alpha should be introduced after Eq. 8 (before the "Second, ...")
    • Eq. 11 has flipped indices for \tau
    • The explanation of how metrics are computed is included in the "Baselines" paragraph, which is quite confusing
    • It is unclear what the forward operator is for the AFM problem.
  • The paper limits comparisons to flow-based methods. It would be nice to at least compare with one or two methods outside this family, to understand whether flow methods are close to optimality or not, for example, with end-to-end methods (e.g., Reconstruct Anything Model, ICLR'26) or with other Bayesian methods (Bayesian Experimental Design Via Contrastive Diffusions, ICLR'25).

Questions for authors
  • Can you give some theoretical intuition on why the reflected noise helps?
  • All the forward operators in the paper seem to be unitary, which simplifies the projection step significantly - would this method work on problems with a fast-decaying spectrum, eg sparse-view tomography?
5 comments
AuthorsOpenReview ↗

W1 + Q1. Why RNC works

Thank you for the remark. We agree that the original presentation did not sufficiently explain the motivation for RNC, we have now introduced two appendix sections (B.4 and B.5) to address this concerns.

With independent re-noising, each iteration introduces a perturbation nullspace. With reflection, successive perturbations push in opposite directions while the prior refines the reconstruction. Although due to the nonlinear denoiser these oscillation do not cancel exactly, we show that when we follow the center of the alternating motion, the effect of the perturbation is cancelled under some stated assumptions.

AuthorsOpenReview ↗

W2. Multiple typos and presentation errors.

Thank you for identifying these mistakes. We have revised the manuscript to address each point.

  • We've filled up the white space after Eq. 12.
  • τ\tau has now the appropriate sub-indices in Eq. 9.
  • α\alpha is now introduce after Eq. 8
  • Eq. 9 has now the indices of τ\tau in the correct order
  • The baselines paragraph now makes no mention of the metrics
  • The forward operator of the AFM problem is now described in the AFM paragraph in section 5.1.
AuthorsOpenReview ↗

W3. Comparison with baselines outside the flow-based family

Thank you for suggesting RAM, and Bayesian Experimental Design via Contrastive Diffusions,which selects measurements that maximize expected information gain using diffusion‑based posterior sampling. In our experiments we keep the acquisition sequence fixed across methods and assess reconstruction quality and computational cost as measurements arrive. Allowing BED to optimise acquisitions would alter both the data and the reconstruction process, while fixing its acquisitions would eliminate its main contribution.Thus BED would be an interesting future direction for complementary work rather than a reconstruction baseline. Combining BED’s measurement‑selection strategy with MPR‑Flow could be a promising extension.

We implement RAM in our sequential-acquisition setting, evaluated in 128 images against MPR-flow, with the same settings as in Table 2.

MethodInpainting 10%Inpainting 30%Random matrixMRI σ=0MRI σ=0.03DNA 30%
RAM13.19 / 0.38330.41 / 0.9139.72 / 0.13228.94 / 0.78533.47 / 0.71740.53 / 0.978
MPR-Flow, N=1024.59 / 0.77831.96 / 0.94125.14 / 0.73536.08 / 0.87835.50 / 0.89341.81 / 0.978
MPR-Flow, N=5025.76 / 0.82632.80 / 0.95328.19 / 0.83736.35 / 0.88435.70 / 0.90342.22 / 0.987

We observe that RAM struggles the most in inpainting 10% and in Random Matrix, but overall tends to land below MPR-flow with a budget of 10 NFEs, however it does remain competitive in Inpainting 30% surpassing the results reported in Table 2 for both Flower and PnP-Flow.

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

This paper proposes MPR-Flow for reconstructing a fixed target from progressively accumulated measurements. It combines acquisition-dependent warm starting (ADWS) with reflected noise coupling (RNC) within the PnP-Flow framework. Experiments cover inpainting, random-matrix sensing, MRI and synthetic DNA AFM reconstruction. The evaluation includes component ablations, runtime measurements and stagewise reconstruction curves.

Critical strengths
  1. The problem setting is meaningful. Reusing previous computations during online acquisition better reflects practical workflows than solving each stage from scratch.
  2. Reusing a pretrained flow prior is a practical advantage.
  3. The paper includes component ablations. Table 2 evaluates ADWS and RNC separately, while the appendices provide stage-by-stage results and some runtime measurements.
  4. The proof of Gaussian noise marginal preservation appears sound under exact orthogonal projection. The appendix also acknowledges covariance errors arising from approximate projections, which appropriately qualifies the guarantee.
Critical weaknesses
  1. The contribution of the ADWS schedule is not isolated. ADWS combines initialization from the previous reconstruction with a later starting time, which reduces re-noising and increases the data-consistency step size for negative α\alpha. Please compare against a tuned fixed-time warm start and warm-started FLOWER, and clarify how these effects interact.

  2. The RNC analysis does not establish the claimed cancellation mechanism. Gaussian noise marginals alone do not guarantee compatibility with the pretrained flow’s input distribution, since the reconstruction and noise are dependent. How do the results in Appendix B explain perturbation reduction after nonlinear reconstruction updates?

  3. The benefit of subspace-dependent coupling remains unclear. Random-matrix experiments use global sign reversal (P=0, R=−IP=0,\ R=-I). A direct comparison between global reversal and measurement-nullspace RNC on inpainting or MRI is needed.

  4. The experimental coupling is insufficiently specified. The qualitative experiments use ρ=0.97\rho=0.97, but this parameter is absent from the main equations and Algorithm 1. What does it control, and does the implementation match the analyzed coupling?

  5. The large random-matrix gains need further explanation. At 10 NFE, PSNR improves from 14.7 to 25.1 dB, with 9.5 dB attributable to ADWS in Table 2. Why is warm starting particularly effective here, and do these gains persist against tuned fixed-time warm starts under matched cumulative budgets?

  6. Final measurement consistency is unclear. Algorithm 1 returns the final prior prediction, which can modify measurement-supported components. Please report the final measurement residuals and assess whether an additional data-consistency step is beneficial.

Questions for authors
  1. Under the same computational budget, how much does MPR-Flow improve over warm-started PnP-Flow/FLOWER with a fixed starting time? 2.What is the precise definition of ρ=0.97\rho=0.97? Does the noise recursion used in the reported experiments match Equation (13)? 3.Why do the random-matrix experiments use global sign reversal? Does measurement-aware projection outperform this simple fallback? 4.Which independent data split is used for hyperparameter tuning? How are the MRI data preprocessed and the forward model constructed, and how is the data-fidelity term defined in the noiseless setting? 5.Are RNC’s additional gains consistent across different random measurement realizations and noise seeds? What is the total online latency for a single sample?
8 comments
AuthorsOpenReview ↗

W1 + Q1. Stronger Warm Start Comparison and extension to FLOWER

We thank the reviewer for the opportunity given to clarify this point. To isolate the contribution of the stage dependent starting-time within ADWS, we compare it with a tuned fixed starting time warm starting strategy, we compare the two approaches, both for Flower and for PnP-flow, under the same conditions stated in Tab. 2:

MethodInpainting 10%Inpainting 30%MRI σ=0\sigma=0MRI σ=0.03\sigma=0.03Random matrix
FLOWER15.487 / 0.359325.011 / 0.737832.239 / 0.844332.094 / 0.849513.753 / 0.2547
FLOWER + fixed warm start21.687 / 0.634230.160 / 0.908435.114 / 0.886634.533 / 0.887021.076 / 0.5995
FLOWER + ADWS22.376 / 0.679330.407 / 0.916135.275 / 0.887834.689 / 0.888521.127 / 0.6011
FLOWER + ADWS + RNC23.090 / 0.714431.306 / 0.929635.804 / 0.869835.360 / 0.889421.643 / 0.6345
PnP-Flow16.290 / 0.391726.113 / 0.794230.563 / 0.654230.591 / 0.671414.653 / 0.3030
PnP-Flow + fixed warm start22.952 / 0.694230.658 / 0.917935.270 / 0.890934.577 / 0.886424.025 / 0.6962
PnP-Flow + ADWS23.634 / 0.733230.843 / 0.923735.403 / 0.891834.704 / 0.887624.203 / 0.7031
PnP-Flow + fixed warm start + RNC23.816 / 0.740531.762 / 0.936335.961 / 0.876935.413 / 0.892524.537 / 0.7163
MPR-Flow24.589 / 0.778531.962 / 0.940936.075 / 0.878035.495 / 0.893325.136 / 0.7347

ADWS brings small but consistent gains across all experiments compared to a fixed warm start. Reconstructions on inpainting 10% and random matrix benefit the most from ADWS (respectively +0.77 and +0.6 PSNR). We also see that the ranking of the ablations transfers to FLOWER, with RNC + ADWS bringing in the most gains among the FLOWER ablations.

As pointed out, we do observe some interaction between α\alpha and tmaxt_{max}. As we increase negative α\alpha, the size of the data fidelity steps increases as shown in Fig. 10. When the starting time is too close to 1, high values for α\alpha might lead to numerical instabilities, as illustrated by the truncated curves.

AuthorsOpenReview ↗

W2. Theoretical justification of RNC

Thank you for the observation. We agree that the Gaussian marginals do not guarantee compatibility with the pretrained flow's input distribution. We added new appendix subsections (B.3 and B.4) where we make use of Tweedie's representation to describe an ideal denoising function associated with an independently corrupted training distribution, and under these settings analyze the effect of reflected noise coupling.

We show that alternating displacements from the noise cancel out, and that the reconstruction converges to a data-consistent local optimum as the iteration budget grows.

AuthorsOpenReview ↗

W3 + Q3: Random-matrix experiments and global sign reversal fallback

Thank you for the remark. To isolate the gains of RNC over global sign reversal, we test both noise formulations. The following results are reported after being averaged over 128 images; the solvers are warm-started with ADWS, and the evaluations are done over 2 solver seeds and 3 different random measurement selections per image. All the solvers use 10 acquisition stages.

Noise ruleInpainting 10%Inpainting 30%MRI σ=0MRI σ=0.03Random matrix
Global flip (fallback)24.010 / 0.747931.397 / 0.932436.505 / 0.889035.728 / 0.900924.896 / 0.7239
RNC24.340 / 0.771131.843 / 0.941036.606 / 0.879235.913 / 0.902225.018 / 0.7392
Difference+0.330 / +0.0232+0.446 / +0.0086+0.101 / −0.0098+0.185 / +0.0013+0.122 / +0.0153

We observe that RNC brings small but consistent gains against global sign reversal. Inpainting benefits most from RNC, while gains on MRI data are more marginal, which might be due to the fact that the measurement is more informative than in the case of inpainting where we see the largest gains.

The reason we were using the fallback for Random matrix is computational, in dense Gaussian sensing the normal operator is not a scaled orthogonal projector, so our fast projector‑based noise decomposition cannot be used. We therefore resort to global sign reversal, given that it also preserved the noise distribution and does yield a benefit in comparison to defaulting the the standard re-noising procedure with independently sampled noise. An exact measurement‑aware decomposition remains possible via P=A∗(AA∗)−1AP = A^{*}(AA^{\ast})^{-1} A , but it incurs linear‑algebra costs that grow with the acquisition operator.

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

This paper investigates measurement-progressive inverse problems, in which a fixed target is reconstructed from a growing, nested set of measurements. The authors propose MPR-Flow, a plug-and-play solver that uses a pretrained flow-matching prior and incorporates two components. Acquisition-dependent warm starting reuses the preceding reconstruction and shifts the starting flow time toward the data endpoint as acquisition progresses. Reflected noise coupling introduces correlation between successive re-embedding noises; in its projector-based form, it resamples noise in measurement-supported directions and reverses the sign of the inherited nullspace component. The authors prove that this construction preserves standard Gaussian noise marginals when the projector is exactly orthogonal. Experiments cover progressive random inpainting, random-matrix sensing, MRI reconstruction using fastMRI brain data, and synthetic DNA AFM reconstruction. The reported results indicate improved reconstruction quality over the evaluated cold-start flow-based baselines, particularly at low per-stage network-evaluation budgets, and reduced computational cost for reaching several baseline quality targets. In the reported ablations, warm starting accounts for most of the improvement, while noise coupling provides additional PSNR gains.

Critical strengths
  1. The paper addresses a useful problem: measurements arrive in stages, and a reconstruction is needed before all the data have been collected. Reusing the previous reconstruction is a practical way to make use of work that has already been done. The proposed changes can be added to a pretrained flow model without further training of that model, which makes the approach relatively easy to adopt.

  2. The experiment results show better reconstruction quality than the tested cold-start baselines on inpainting, random-matrix sensing, MRI, and synthetic DNA AFM data. The gains are particularly clear when only a few network evaluations are allowed at each stage. The timing results also show that the method can match baseline reconstruction quality with less computation in several settings.

Critical weaknesses

1.A stronger warm-start comparison is missing. Table 2 shows that ADWS accounts for most of the improvement. What is less clear is how much comes from reusing the previous reconstruction and how much comes from changing the starting time across stages. I would like to see a comparison with a fixed, tuned, nonzero starting time, using the same previous reconstruction and budget. This matters given earlier work on informative initialization and sequential warm starts, including CCDF [1] and SeqDiff [2]. Appendix A.2 also describes a warm-start version of FLOWER, but I could not find results for it. That comparison would make the advantage over existing solvers easier to judge. 2.The role of the measurement subspace is still unclear. In the random-matrix experiment, RNC uses P = 0 and R = −I, so it becomes global sign alternation without using the measurement subspace. The gains in this case do not tell us what the projector-based coupling adds. A useful comparison would be independent noise, global sign alternation, and nullspace coupling on the same task, with the same warm start and budget. Antithetic noise methods [3] are relevant here, although their coupling scheme differs from RNC. The Gaussian-marginal proof is useful, but it does not explain how these correlations affect reconstruction error after the nonlinear denoising step. 3. The AFM numbers need to be reconciled. I had trouble matching the AFM results across the tables and figures. The final PnP-Flow PSNR in Fig. 16 appears to be about 35.3 dB, while Table 4 reports 36.69 dB. Table 4 also says the results use 10 acquisition stages, whereas Section 5.1 and Fig. 16 use 30. These may be different experimental settings, but that needs to be stated. It is also unclear which baseline value supports the stage-13 matching result in Table 6. Please identify the settings behind each result and make the comparison used to claim measurement savings explicit.

  1. The implemented data step needs to be spelled out. The definition F_s(x) = ‖A_s x − y_s‖² / (2σ_s²) is undefined for the noiseless setting σ_s = 0, and it produces a large inverse-variance factor when the noise is small. Does the code use a variance floor, a normalized loss, or an extra step-size factor? The negative values of α in Table 5 also need an explanation: with η(τ) = (1 − τ)^α, the step grows as τ approaches one, whereas the original PnP-Flow uses positive exponents [4]. This could be an intentional change, but the equations should match the implementation. Figure 10 should also mark the divergent settings mentioned in Appendix F.2.

  2. The parameter ρ is not defined. Figures 6–9 list ρ = 0.97, but this parameter does not appear in Eq. (13), Algorithm 1, or the hyperparameter table. What does it control? If it changes the noise coupling, the paper needs to give the actual update and explain how it relates to the update analyzed in Appendix B. Please also state whether the main quantitative results use this parameter. As written, I cannot tell whether the figures and the main tables use the same noise update.

  3. The projector check depends on the scale of A. In Appendix B.3, rescaling A to cA changes the residual in Eq. (29) to c²δ, but leaves both X = H/λ̂ and the covariance defect 2(X² − X) unchanged. With a fixed absolute tolerance, a small screening residual therefore need not mean a small covariance error. Was the operator normalized before this check, or was the tolerance adjusted to its scale? The probe count and acceptance tolerance should be reported, together with an explanation of the covariance accuracy over the chosen iteration budget. This concern applies to the approximate-projector extension.

References [1] H. Chung, B. Sim, and J. C. Ye. “Come-Closer-Diffuse-Faster: Accelerating Conditional Diffusion Models for Inverse Problems Through Stochastic Contraction.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 12413–12422, 2022. [2] T. S. W. Stevens, O. Nolan, J.-L. Robert, and R. J. G. van Sloun. “Sequential Posterior Sampling with Diffusion Models.” IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5, 2025. [3] J. Jia, S. Liu, B. Song, W. Yuan, L. Shen, and G. Wang. “Antithetic Noise in Diffusion Models.” International Conference on Learning Representations (ICLR), 2026. [4] S. Martin, A. Gagneux, P. Hagemann, and G. Steidl. “PnP-Flow: Plug-and-Play Image Restoration with Flow Matching.” International Conference on Learning Representations (ICLR), 2025.

Questions for authors

See Weaknesses.

8 comments
AuthorsOpenReview ↗

W1. Warm Start Comparisons and Flower Implementation.

We thank the reviewer for the suggestion; we agree that reusing the previous reconstruction is an important source of improvement. To further illustrate the gains achieved by ADWS, we compared its performance against MPR-Flow using a fixed warm start. We present the results in the following table. Both approaches use N=10 for all runs, with hyperparameters tuned separately, and final results averaged over 128 images. ADWS brings small but consistent gains across all experiments compared to a fixed warm start. Reconstructions on inpainting 10% and random matrix benefit the most from ADWS (respectively +0.77 and +0.6 PSNR), while ADWS and fixed warm start are on par on MRI data.

MetricInpainting 10%Inpainting 30%Random matrixMRI σ=0.0\sigma=0.0MRI σ=0.03\sigma=0.03
Fixed WS (PSNR / SSIM)23.816 / 0.740531.762 / 0.936324.537 / 0.716335.961 / 0.876935.413 / 0.8925
ADWS (PSNR / SSIM)24.589 / 0.778531.962 / 0.940925.136 / 0.734736.075 / 0.878035.495 / 0.8933
PSNR gain (mean ± SD)+0.7729 ± 0.6695+0.2001 ± 0.1958+0.5999 ± 0.3176+0.1136 ± 0.2325+0.0819 ± 0.1484
SSIM gain (mean ± SD)+0.03795 ± 0.02135+0.00466 ± 0.00290+0.01840 ± 0.01651+0.00113 ± 0.00249+0.00082 ± 0.00179

Comparison with CCDF and SeqDiff:

Regarding the comparison with earlier work on informative initialization, We implemented comparisons with CCDF and SeqDiff. However, as no official implementation exists we adapted our own implementation from efzero/diffusioncontroller following the measurement-consistency projections of the original paper. In addition, we also evaluated SeqDiff using our own re-implementation. Both use a pretrained CelebA-HQ score model [1]. For informed initialization, we also tested DeepFill v2 [2], as in the original CCDF inpainting experiments. We fine-tuned these methods and compared the methods on the same images, measurements, and ten-stage sequential protocol used for MPR-Flow.

We evaluate CCDF and SeqDiff at 10, 50, and 100 NFEs.

For this comparison, we selected publicly available pretrained models with matching NCSN++ architectures: a rectified-flow model [3], and a score-based model from [1], both trained on CelebA-HQ: a rectified-flow model for MPR-Flow and a score-based model for CCDF and SeqDiff. The table reports PSNR / SSIM

MethodNFEs/stageInpainting 10%Inpainting 30%
CCDF1019.212 / 0.501123.146 / 0.5678
SeqDiff1015.125 / 0.462316.566 / 0.4982
CCDF5021.264 / 0.534426.020 / 0.7693
SeqDiff5023.538 / 0.689328.912 / 0.8500
CCDF10021.220 / 0.574926.047 / 0.7891
SeqDiff10025.494 / 0.760330.071 / 0.8865
MPR-Flow1026.195 / 0.769230.917 / 0.8970
MPR-Flow5027.321 / 0.809131.467 / 0.9190

MPR-flow achieves the highest PSNR and SSIM at both budgets, at 100 NFE, SeqDiff reaches 25.494 dB / 0.7603 SSIM for 10% inpainting and 30.071 dB / 0.8865 for 30% inpainting, compared with 26.195 dB / 0.7692 and 30.917 dB / 0.8970 for MPR-Flow at 10 NFE.

In addition, we included the results achieved by the adaptation of Flower as suggested. The results are averaged over 128 images, with a budget of 10 NFEs.

N=10

MethodInpainting 10%Inpainting 30%Random matrixMRI σ=0.0\sigma=0.0MRI σ=0.03\sigma=0.03
FLOWER15.487 / 0.359325.011 / 0.737813.753 / 0.254732.239 / 0.844332.094 / 0.8495
FLOWER + warm start21.687 / 0.634230.160 / 0.908421.076 / 0.599535.114 / 0.886634.533 / 0.8870
FLOWER + ADWS22.376 / 0.679330.407 / 0.916121.127 / 0.601135.275 / 0.887834.689 / 0.8885
FLOWER + ADWS + RNC23.090 / 0.714431.306 / 0.929621.643 / 0.634535.804 / 0.869835.360 / 0.8894
MPR-Flow24.589 / 0.778531.962 / 0.940925.136 / 0.734736.075 / 0.878035.495 / 0.8933

We observe in the results that warm starting improves Flower for all tasks. ADWS provides further improvement, and both ADWS and RCN provide positive improvements in all problems in PSNR. MPR-Flow still achieves the highest metrics when compared against Flower even with the incorporation of the ablations.

AuthorsOpenReview ↗

References

[1] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. “Score-Based Generative Modeling through Stochastic Differential Equations.” ICLR, 2021.

[2] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S. Huang. “Free-Form Image Inpainting with Gated Convolution.” ICCV, pp. 4471–4480, 2019. (DeepFill v2.)

[3] Xingchao Liu, Chengyue Gong, and Qiang Liu. “Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow.” ICLR, 2023.

AuthorsOpenReview ↗

W2. Role of measurement subspace

We thank the reviewer for his careful remarks. Indeed, RNC is more suited for reconstruction than global sign alteration (fallback), which we highlight both empirically and theoretically. First, to further extend the theoretical guarantees of our approach, we show that the error term related to the sampled noise is cancelled out through RNC; see Appendix B.5 of the revised manuscript for the proof. Second, we also present empirical results for different noise formulations in the following table, where the solvers are warm-started with ADWS.

Random Split reflects the noise of random dimensions in a proportion equal to RNC. iid incorporates independently distributed noise in all dimentions, aligning with the renoising step of PnP-Flow.

Noise ruleInpainting 10%Inpainting 30%MRI σ=0MRI σ=0.03Random matrix
iid23.318 / 0.729830.637 / 0.922935.945 / 0.898235.066 / 0.893523.969 / 0.6932
Random split23.986 / 0.747731.234 / 0.930336.374 / 0.891335.568 / 0.899324.878 / 0.7238
Global flip (Fallback)24.010 / 0.747931.397 / 0.932436.505 / 0.889035.728 / 0.900924.896 / 0.7239
RNC24.340 / 0.771131.843 / 0.941036.606 / 0.879235.913 / 0.902225.018 / 0.7392

Although global noise flipping improves over independent noise, RNC surpasses the fallback in PSNR and SSIM in every case. These results show that using the measurement subspace provides additional benefit beyond global sign reflection. This conclusion is in agreement with the extended theoretical justification we added to the paper. Furthermore, if obtaining this measurement subspace is costly, we can fall back to global noise flipping and still see a benefit compared to the initial PnP-Flow IID noise sampling scheme.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Authors’ Answers and Additional Experiments

Global Comment: We thank the reviewers for their careful feedback. Below we summarize the strengths they identified and our answers to their shared concerns. Details can be found in the individual responses, and the changes are highlighted in blue in the revised paper.

Strenghts Reviewers generally measurement progressive reconstruction a useful problem (xWWe, ucvP, Gyin), agreeing that reusing previous reconstructions is practical (xWWe, ucvP). Reviewers recognized the reconstruction quality gains over cold-started baselines across operators (xWWe, ucvP, Gyin, 8tww), and also identified the computational efficiency gains (xWWe, 8tww)

Main Concerns

1. Separating acquisition-dependent time scheduling from reconstruction reuse (xWWe, ucvP, Gyin)

We added a tuned fixed-start control under matched reconstruction budgets. Comparisons show that most of the improvement from ADWS comes from the reuse of the previous reconstruction, and the acquisition dependent provides a smaller consistent increase in the gain.

2. Theoretical Motivation from RNC (xWWe, ucvP, Gyin)

We added Appendices B.4-B.5 to analyze the effect of RNC in the reconstruction dynamics. Following the midpoint of the oscillations caused by reflecting the noise we found that the cumulative noise residual is controlled by the flow time.

3. Expansion to FLOWER (xWWe, ucvP, 8tww)

We incorporated the results for FLOWER with fixed warm start, ADWS, and ADWS combined with RNC. The reconstruction reuse greatly improved results for the measurement progressive problem setting, going from (ADWS added from 2.6 dB to 7.37 dB, and ADWS+RNC added from 3.27 dB to 7.89 dB).

Summary of revisions (Highlighted in blue in the updated paper)

  • Sec 4.2: Sentence clarifying the role of α\alpha, fixed typos in equations (9) and (11), and in the definition of Fs(x)F_s(x), added sentence referencing newly added proofs.
  • Fig. 3: Fixed typo in caption
  • Sec 5.1 and Tab. 1: Clarified operator for AFM, added new baseline.
  • Sec 5.3 and Tab. 3: Added extra ablation comparing fixed warm starting with ADWS
  • App. A. 2: Added metrics for ADWS and RNC applied to Flower.
  • App. B. 4 and B.5: Added theoretical developments regarding the effects of RNC.
  • App. E and Figs. 6, 7, 8, and 9: Fixed the not updated captions.
  • Fig. 10: Added clarification to caption
  • App. G, Fig. 16 and Tab. 8: Changed Fig. 16 and Tab. 4. after running experiments for different parameters.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.