List of Papers Browse by Subject Areas Author List
Abstract
Zero-shot MRI reconstruction relies on generative priors, but single-modality unconditional priors produce hallucinations under severe ill-posedness.
In many clinical workflows, complementary MRI acquisitions (e.g. high-quality structural scans) are routinely available, yet existing reconstruction methods lack mechanisms to leverage this additional information.
We propose MPFlow, a zero-shot multi-modal reconstruction framework built on rectified flow that incorporates auxiliary MRI modalities at inference time without retraining the generative prior to improve anatomical fidelity.
Cross-modal guidance is enabled by our proposed self-supervised pretraining strategy, Patch-level Multi-modal MR Image Pretraining (PAMRI), which learns shared representations across modalities.
Sampling is jointly guided by data consistency and cross-modal feature alignment using pre-trained PAMRI, systematically suppressing intrinsic and extrinsic hallucinations.
Extensive experiments on HCP and BraTS show that MPFlow matches diffusion baselines on image quality using only 20% of sampling steps while reducing tumor hallucinations by more than 15% (segmentation dice score). This demonstrates that cross-modal guidance enables more reliable and efficient zero-shot MRI reconstruction. Code is available at https://github.com/edshkim98/MPFlow.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2812_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/edshkim98/MPFlow
Link to the Dataset(s)
N/A
BibTex
@InProceedings{KimSeu_MPFlow_MICCAI2026,
author = { Kim, Seunghoi AND Jin, Chen AND Tregidgo, Henry F. J. AND Figini, Matteo AND Alexander, Daniel C.},
title = { { MPFlow: Multi-modal Posterior-Guided Flow Matching for Zero-Shot MRI Reconstruction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16888},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposed a flow matching-based posterior sampling method that leverages multi-modal information learned from contrastive learning. Experiments demonstrated the effectiveness of the proposed component.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The idea of introducing a multi-modal feature-level constraint to reduce extrinsic hallucination is interesting.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The compared baselines are limited to diffusion models. Recent work on flow-based inverse problems solving is missing. For example, “PnP-flow: Plug-and-play image restoration with flow matching”, ICLR 2025.Also, for diffusion-based baselines, the paper only includes DPS related approaches. Better performing solutions like DDNM (Zero-Shot Image Restoration Using Denoising Diffusion Null-Space Model, ICLR 2023) were omitted. 2.The evaluation on MRI reconstruction is limited to simulated magnitude data from the BraTS dataset. Real-world complex-value multi-coil raw data were not evaluated. 3.The initial noise optimization module did not lead to significant improvement according to the ablation study, which however increases computational cost.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The weaknesses outweigh the strengths significantly. I would lean to reject the submission.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The strongest unresolved issue is the MRI realism/generalizability concern. The paper frames the method as MRI reconstruction and defines the forward model broadly enough to include complex-valued k-space tasks, but the actual evaluation is on HCP super-resolution and BraTS simulated accelerated reconstruction using magnitude-style image data, with no real complex-valued multi-coil raw MRI experiment. The paper’s own experimental setup lists HCP T2 super-resolution and BraTS FLAIR k-space reconstruction with T1 auxiliary images, rather than fastMRI-like raw multi-coil reconstruction. This limits the strength of the clinical reconstruction claim. This is a major limitation, not a minor extension. Since the method is presented as an MRI reconstruction framework, the absence of real multi-coil raw-data experiments substantially weakens the empirical validation.
Review #2
- Please describe the contribution of the paper
The paper makes a strong and well-motivated contribution to zero-shot MRI reconstruction. Its central idea—integrating auxiliary anatomical information into a posterior-guided generative reconstruction framework—is both elegant and conceptually insightful. In particular, the proposed method leverages auxiliary MR images at inference time, without requiring retraining or modification of the generative prior itself.
The patch-level multimodal representation learning framework(PAMRI) is practical and technically well aligned with the nature of MRI reconstruction. By operating in a learned latent space rather than enforcing direct pixel-level consistency between different MR contrasts, the method offers an effective way to handle the intensity and contrast mismatch across modalities. The use of mutual-information-based adaptive weighting in the patch alignment objective is particularly thoughtful and strengthens the multimodal formulation.
The authors validate the method across two distinct reconstruction tasks and datasets, covering both super-resolution and k-space reconstruction settings. In addition to standard image-quality metrics, they evaluate hallucination-related behavior and downstream task performance, which makes the empirical validation substantially more convincing. Overall, the paper presents a clear methodological contribution supported by comprehensive experiments.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Rather than retraining or redesigning the generative prior to incorporate auxiliary inputs, the method uses the auxiliary modality to guide posterior sampling, which is a clean and flexible design.
2.The proposed PAMRI module is a particularly strong technical component. The use of patch-level multimodal representation learning is well motivated for MRI reconstruction, where preserving local anatomical structure is essential. Moreover, aligning modalities in a learned latent space, instead of imposing pixel-level(may be TV/Sobel) consistency, is especially suitable for multi-contrast MRI. In this regard, the adaptive mutual-information-based weighting is also a thoughtful addition: it provides a practical mechanism to account for cross-contrast mismatch while still extracting shared anatomical information. This is an original and technically meaningful way to exploit auxiliary MR images.
3.The paper is very clearly written.
4.The evaluation is comprehensive and convincing. The method is validated on two distinct MRI reconstruction settings, which strengthens the claim that the framework is general. In addition, the authors also assess hallucination-related behavior and downstream clinical relevance through segmentation-based analysis. It demonstrates that the method improves not only reconstruction fidelity in a numerical sense, but also structural and semantic reliability in ways that are meaningful for medical imaging applications.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The paper is overall strong, and I did not identify major conceptual flaws that substantially weaken the main claims.
One minor concern is the behavior of the DIP baseline, which appears visibly noisy and may be operating in an overfitting regime. Since DIP is known to be sensitive to stopping criteria, it would be preferable to clarify whether early stopping or a tuned stopping protocol was used for this baseline. A better-controlled DIP comparison would improve the fairness of the experimental presentation.
That said, this issue does not materially affect my confidence in the proposed method or the main conclusions of the paper, since the gains over the stronger generative baselines remain convincing.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My recommendation is primarily based on the clarity of the paper, the originality and elegance of the core idea, and the strength of the experimental validation. The method is conceptually well designed, and the writing makes the technical motivation and implementation easy to follow.
In addition, the proposed latent-space multimodal alignment strategy is well suited to MRI, where direct intensity correspondence across contrasts is often unreliable. The experiments are comprehensive, spanning multiple datasets, tasks, and evaluation protocols, including both conventional reconstruction metrics and downstream assessments related to hallucination and clinical relevance. Taken together, these factors make the work a high-quality and compelling contribution.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The primary idea of this paper, auxiliary mri data informed reconstruction is compelling, and the authors have effectively implemented it using the modern deep learning concept of latent space alignment. I believe this paper could have a positive impact on future research in MRI reconstruction, considering that clinical acquisitions typically do not rely on single-modality data.
Review #3
- Please describe the contribution of the paper
This paper proposes MPFlow, a framework for zero-shot MRI reconstruction that introduces an auxiliary MRI modality as inference-time posterior guidance without retraining the unconditional generative prior. The method consists of two components: (1) PAMRI, a self-supervised patch-level cross-modal pretraining module that learns shared representations between target and auxiliary modalities; and (2) a multi-modal posterior-guided flow matching procedure that injects both data-consistency and PAMRI feature-alignment gradients into a rectified flow sampler, together with an initial noise selection mechanism. Experiments are conducted on HCP T2 super-resolution and BraTS FLAIR accelerated reconstruction, using fully-sampled T1 as the auxiliary modality. The authors report improvements in image quality metrics (PSNR, SSIM, LPIPS) and in hallucination-related indicators (SHAFE and tumor segmentation Dice).
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Meaningful and moderately novel problem formulation. Most zero-shot reconstruction methods rely on a unimodal unconditional prior. Framing inference-time multimodal guidance as a way to reduce hallucination—without retraining the prior—is a practically motivated and reasonably original angle.
2.Coherent end-to-end design. The pipeline from cross-modal representation learning (PAMRI) to posterior guidance and initial noise selection tells a consistent story; the components are not arbitrarily concatenated.
3.Evaluation beyond PSNR/SSIM. Including SHAFE and tumor segmentation Dice as proxy hallucination metrics is a meaningful step toward clinically relevant evaluation, going beyond standard pixel-level distortion measures.
4.Efficiency argument is concretely demonstrated. The authors show that fewer sampling steps can still match or exceed diffusion-based baselines, which is a substantive efficiency claim rather than a theoretical one.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.FlowDPS comparison is missing. All main comparisons are against diffusion or DIP methods, despite the prior being a rectified flow. The paper cites FlowDPS [13] but provides no experimental comparison. Since FlowDPS is a directly comparable flow-based posterior sampling method, a comparison would calibrate how much gain comes from the multimodal guidance over a competitive unimodal flow baseline, rather than only over the paper’s own base model in Table 3a.
2.PAMRI necessity is not sufficiently justified. The ablation only compares w/ PAMRI vs. w/o PAMRI, without testing simpler auxiliary guidance alternatives. It therefore remains unclear whether the full patch-level contrastive design is necessary over a simpler baseline.
3.Hidden computational overhead makes Table 1 comparisons unfair. The initial noise selection (S=8, t_noise=0.2T) incurs 8x20+80=240 NFE at the nominal T=100 setting—2.4× the baseline cost. Efficiency comparisons must be re-reported at matched NFE.
4.Patch operator P(⋅) undefined; background handling unspecified. The sampling strategy for P(⋅) is never stated. More critically, without background masking on skull-stripped BraTS data, all-zero patches will cause NMI to be undefined and contrastive representation collapse. Both issues must be addressed for reproducibility.
5.Symbol clash in Eq. (6). z is used both for the initial noise z ~ N(0, I) in Eq. (4) and for the contrastive features z_i ∈ {u_i, w_i} in Eq. (6). Additionally, the definition of z_j in a positive pair is never stated explicitly. The notation should be disambiguated.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
1.The abstract and contribution list should explicitly state that PAMRI requires pretraining on paired, registered multi-modal data from the target domain; the system is not a plug-and-play extension of any off-the-shelf prior. 2.Clarify whether PAMRI and the generative prior share the same training split and dataset. 3.Multi-auxiliary scalability is unaddressed. Eq. (2) naturally extends to multiple auxiliary modalities, yet PAMRI’s single ψ(⋅) encoder and the guidance term in Eq. (8) are designed for one auxiliary input only. The authors should discuss how the framework would extend and how overlapping information across modalities would be handled. 4.Clinical applicability boundary. Both datasets use pre-registered data, so the registration assumption is never stress-tested. It would be worth discussing how the method is expected to behave under imperfect registration (e.g., differing slice thickness, inter-session motion), and whether registering a low-quality target to a high-quality auxiliary is a practical prerequisite in real workflows.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The core idea is genuinely interesting and the method design is coherent. The hallucination evaluation via SHAFE and Dice is reasonable for a venue like MICCAI, though the paper would benefit from a brief acknowledgment of their limitations as proxy metrics. The main concerns driving the score are: (W1) FlowDPS baseline is absent, leaving the source of gains unattributed; and (W3) the true NFE at T=100 is 240, making Table 1 comparisons materially unfair—this requires re-running at matched NFE, not just clarification. A satisfactory rebuttal on these two points would increase confidence in the score.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The submitted paper was already close to acceptance based on a coherent multimodal posterior-guided reconstruction framework and solid hallucination-related and downstream-task evaluation. The rebuttal clarifies my main concerns about the FlowDPS-style comparison and NFE accounting, but much of the supporting evidence is rebuttal-only and should be treated as confidence-increasing rather than equivalent to submitted experiments. Overall, the concerns are sufficiently reduced for an Accept recommendation, but not for highlight consideration.
Author Feedback
We thank all reviewers for their constructive feedback and for recognizing the importance of mitigating hallucinations in MRI reconstruction. We are pleased that the reviewers found our motivation strong (R2, R3), our method well-designed (R1-R3), and our experiments thorough (R2, R3). Missing baselines [R1.7.1, R3.7.1]: We evaluated FlowDPS (ICCV’25) and DDNM (ICLR’23) on HCP at T=100 using our pretrained rectified flow and diffusion priors. FlowDPS: SSIM 0.624±0.049, SHAFE 0.216±0.043; DDNM: SSIM 0.764±0.123, SHAFE 0.221±0.107.Both underperform MPFlow. Adding PAMRI guidance to FlowDPS yields SSIM 0.650±0.047, SHAFE 0.204±0.039, confirming PAMRI’s contribution is sampler-agnostic and complementary to recent posterior sampling methods. Matched-NFE evaluation [R3.7.3]: We agree noise optimization adds NFEs. For fair comparison, we re-ran DPS and DynamicDPS at T=240 (matched to MPFlow’s effective NFE). DPS: SSIM 0.713, SHAFE 0.220; DynamicDPS: SSIM 0.702, SHAFE 0.269.Both remain inferior to MPFlow (SSIM 0.78, SHAFE 0.20), demonstrating that MPFlow’s efficiency advantage is not an artifact of NFE accounting. Removing noise optimization from MPFlow (NFE=100) still yields substantial improvements over Tab. 1 baselines, as shown in Tab. 3.Patch operator [R3.7.4]: P(·) constructs a binary brain mask from the auxiliary image, partitions the image into non-overlapping patches, and selects only patches with sufficient brain content. We will clarify this in the revised manuscript. PAMRI necessity [R3.7.2]: We compared PAMRI against three intensity-based alternatives: Normalized Mutual Information (NMI), Canny edge loss, and pixel-MSE. NMI: SSIM 0.762±0.036, SHAFE 0.210±0.059; edge loss: SSIM 0.745±0.044, SHAFE 0.255±0.155; pixel-MSE: SSIM 0.748±0.037, SHAFE 0.215±0.053.PAMRI outperforms all three. While NMI is the strongest classical baseline, it still operates at the pixel level and cannot capture the fine-grained structural correspondence that PAMRI’s learned latent space provides. We also observed worsening in SHAFE score across all variants compared to using PAMRI (Tab. 3a), indicating that naive intensity-based guidance cannot suppress hallucinations as effectively. Noise optimization [R1.7.3]: While the ablation gain in Tab. 3a appears modest, isolating noise optimization against the vanilla baseline yields SSIM 0.754, SHAFE 0.218, a consistent ~2% improvement. Multi-coil data [R1.7.2]: Our evaluation on magnitude images follows standard practice in recent MRI reconstruction work (DynamicDPS, DiffDeuR); all baselines operate on magnitude data for direct comparability. PAMRI’s cross-modal guidance is orthogonal to magnitude vs. complex domain, the framework extends naturally to complex-valued reconstruction by replacing the forward operator. We agree validation on multi-coil raw data (e.g., fastMRI) is valuable future work. DIP baseline [R2.7.1]: We used measurement-loss-based early stopping for DIP. We tried multiple stopping points and observed minimal change in noise level, likely because measurement loss correlates poorly with image quality on heavily ill-posed problems, making the optimal window narrow and difficult to tune reliably. Notation [R3.7.5]: We will disambiguate the overloaded use of z in Eq. 4 vs. Eq. 6 and clarify positive pair definitions in the revised manuscript. Contribution list [R3.12.1]: We will revise the abstract and contribution list to explicitly state PAMRI’s pretraining requirements. Training data [R3.12.2]: PAMRI and the generative prior use the same training split and dataset, disjoint from the test set. Multi-auxiliary and registration [R3.12.3-4]: We agree multi-auxiliary scalability and stress-testing under imperfect registration are important next steps for clinical deployment.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The reviewers are mixed but see clear promise in the work, with scores of 3, 5, and 4.The paper proposes an interesting multimodal posterior-guided flow matching framework for zero-shot MRI reconstruction, using auxiliary MR images to guide sampling without retraining the generative prior. The PAMRI representation learning component and the use of hallucination-related/downstream metrics are viewed as meaningful strengths.
However, several issues need clarification before a final decision. The rebuttal should address the lack of direct comparison with flow-based posterior sampling baselines such as FlowDPS or other recent flow/inverse-problem methods, since this is important for attributing gains to the proposed multimodal guidance. The authors should also correct the efficiency comparison by reporting matched-NFE results, as the initial noise selection appears to add substantial hidden sampling cost.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
I recommend acceptance. The paper proposes a coherent and interesting multimodal posterior-guided flow matching framework for zero-shot MRI reconstruction, using an auxiliary MR modality to guide sampling without retraining the generative prior. The core idea is well motivated, and the use of PAMRI for latent-space cross-modal alignment is a meaningful way to exploit shared anatomy while avoiding direct pixel-level matching across contrasts.
The rebuttal addressed the main concerns raised during review. In particular, the authors clarified the FlowDPS/DDNM comparisons, matched-NFE accounting, the patch operator and masking strategy, the necessity of PAMRI compared with simpler guidance alternatives, and the role of the noise-optimization module. Reviewers #2 and #3 both support acceptance after rebuttal. Reviewer #1 remains concerned about the lack of real complex-valued multi-coil raw-data validation, which is a valid limitation, but I view it as a scope limitation rather than a fatal flaw for the current contribution.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The authors present an interesting methodology for performing posterior-guided image refinement using an auxiliary image. This wording is specific, as this is not an MRI reconstruction paper, contrary to what the title implies (pointed out by Reviewer 1). However, the authors do define “reconstruction” more broadly as the recreation of high-quality data from limited data. This, however, highlights an issue that is called out in the initial reviews and not addressed. This is the limitation of registered (and seemingly skull-stripped) images in an auxiliary-assisted challenge. This is a strong limitation that is commonly validated in other auxiliary-assisted refinement methods. Further, it takes the method further from the purpose of “reconstruction”, where the assumptions about data preprocessing are necessarily minimal. This is especially true for the two tasks presented here, where registration and skull-stripping would drastically affect the resolution and k-space components.
Additionally, the authors included new experiments conducted in the rebuttal and presented the results as responses to the reviewers’ concerns. This is not allowed under the rebuttal guidelines and cannot carry weight in the acceptance decision at this stage. Unfortunately, most of what the authors present as responses to the author’s concerns fall under this category and must be excluded.
Based on the lack of clarification for some concerns after removing additional experiments, I can only recommend rejection for this paper.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper proposes MPFlow, a zero-shot MRI reconstruction framework that introduces an auxiliary MR modality as inference-time posterior guidance without retraining the unconditional generative prior, combining a patch-level cross-modal representation module (PAMRI) with a multi-modal posterior-guided flow matching sampler. The reviewers agree that the core idea of inference-time multimodal guidance to reduce hallucination is original and practically motivated, that the PAMRI latent-space alignment is well suited to multi-contrast MRI where pixel-level correspondence is unreliable, and that the evaluation is convincing in going beyond fidelity metrics to include hallucination-related and downstream segmentation measures, all of which are established within the submitted manuscript. The rebuttal clarified the effective sampling cost and the efficiency comparison, the positioning relative to flow-based posterior-sampling methods, the patch operator and background handling, and the notation, which together resolved the principal points of concern. The main remaining limitation is that the evaluation rests on magnitude-domain super-resolution and simulated accelerated reconstruction, which the authors note follows standard practice in recent reconstruction work; they frame validation on complex-valued multi-coil data as future work and observe that the framework extends to that setting by replacing the forward operator. The authors are encouraged to state the PAMRI pretraining requirement and the registration assumption explicitly, to qualify the reconstruction claim in line with the magnitude-domain evaluation, and to incorporate the noted clarifications in the camera-ready version.