List of Papers Browse by Subject Areas Author List
Abstract
Recent advances in diffusion and flow models have improved the quality and efficiency of image translation. However, their commonly cited advantages, such as diversity sampling and iterative denoising, may not be considered in the medical domain, where fidelity and quantitative accuracy are paramount. Returning to the essence, the diffusion and flow frameworks cast the distribution transport problem as a continuous-time process, injecting the time variable t as a condition into both the network and the input image. In this paper, we rethink the role of this time-conditioned mechanism in medical image translation and remove other diffusion components in training and inference. Our clean setup yields a powerful one-step deterministic generator. In this context, diffusion is conceptually nothing more than a “Just image Regularizer” with t, or what we call JiR. We support this view with task-driven redesigns of the forward process and the prediction space, and introduce a time-consistency loss that aligns the prediction at any time point with the prediction at the end point. Extensive experiments on cross-modality MR-to-PET and multi-contrast MR translation tasks demonstrate that JiR significantly alleviates overfitting and delivers marked improvements over strong baselines, making diffusion more applicable in this field. We hope our findings motivate the community to revisit the foundations of diffusion and flow models and adopt them more objectively.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0293_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/siyuan-mei/JiR
Link to the Dataset(s)
ADNI: https://adni.loni.usc.edu/
BraTS: https://www.synapse.org/Synapse:syn51156910
BibTex
@InProceedings{MeiSiy_Time_MICCAI2026,
author = { Mei, Siyuan AND Zhang, Yanteng AND Xia, Yan AND Lan, Qizhen AND Sun, Yipeng AND Bayer, Siming AND Li, Zirong AND Ye, Chengze AND Liu, Daiqi AND Jiang, Xiaoqian AND Fan, Fuxin AND Huang, Yixing AND Maier, Andreas},
title = { { Time Matters: Rethinking Diffusion and Flow Models in One-Step Medical Image Translation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces the “Just Image Regularizer” (JiR), exploring time regularization for diffusion and flow matching models in medical image-to-image translation. The authors propose a bridge-image flow, define the network prediction directly in image space, and introduce a time-consistency loss. These modifications yield competitive results for paired image-to-image translation. The method is evaluated on cross-modality MR-to-PET and multi-contrast MR translation, along with an ablation study.
- 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.Overall, the paper is well written and clearly explained. 2.The range of comparison methods and datasets is extensive. 3.The ablation study in Figure 2 is well visualized and provides nice insights into the proposed components. 4.The proposed modifications are flexible and could be adapted to other architectures and settings, making them broadly applicable. 5.Overall this paper is carefully evaluated, with three contributions that improve paired image-to-image translation. Out of curiosity: Is there a possibility to extend this to unpaired image-to-image translation? This would be a nice outlook.
- 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.In Equation 6, how was ω(t) chosen? Is this based on prior work or empirical tuning (e. g. , grid search)? Please clarify. 2.In Figure 2, it is unclear what is meant by the number of sampling steps. Which parts of the training and sampling procedures (Sections 3.2 and 3.3) are affected by this parameter? Additionally, is there an explanation for why BiF performance degrades with more sampling steps? 3.It would be helpful to include information on computational cost (speed and memory) compared to other methods. Is the approach more efficient? A short discussion would strengthen the paper. 4.In Figure 3, some limitations may come from the task itself rather than the competing methods. For example, when generating PET or contrast-enhanced MRI images, the input may not contain all necessary information, making multiple outputs plausible. This ambiguity should be acknowledged in the discussion of results.
- 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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.
(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 paper presents a well-motivated and clearly explained set of modifications for diffusion and flow matching models, supported by experiments and ablations on multiple datasets, providing a useful contribution to image-to-image translation.. Some methodological choices (e.g., the design of ω(t)), technical aspects (number of sampling steps) and practical aspects (efficiency, uncertainty in the task itself) should be addressed in the rebuttal.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The authors identified the optimal setup for one-step generation by analyzing key components, flow interpolants (NiF, PiF, BiF), prediction spaces (v-pred vs. x-pred), and sampling trajectories . The study provides empirical evidence that these frameworks, when stripped of multi-step stochasticity, act as powerful deterministic regularizers that outperform regression and GAN-based baselines.
- 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.While recent work [1] argues that regression models are superior for fidelity-driven MIT due to the hallucinations caused by noise, this work proposes a compelling one-step alternative. The authors clearly articulate that the coupling of noise and time-step t during training is not merely a form of data augmentation but a structural bias that regularizes the learning process across a continuous-time trajectory.
2.The proposed one-step inference (JiR) not only reduces computational overhead but also actively suppresses the hallucinations and noise replication often found in multi-step diffusion.
3.The comparison between different flow types and prediction spaces (especially the advantage of x-pred [2] in manifold alignment) offers significant insights for the medical imaging community.
[1] Rassmann, Sebastian, et al. “Regression is all you need for medical image translation.” IEEE Transactions on Medical Imaging (2026). [2] Li, Tianhong, and Kaiming He. “Back to basics: Let denoising generative models denoise.” arXiv preprint arXiv:2511.13720 (2025).
- 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.Lack of intuitive interpretation for one-step translation. The authors claim that supervising a continuum of intermediate states enhances generalization and prevents shortcut memorization. However, a more intuitive or theoretical explanation of how this continuous trajectory modeling specifically benefits a one-step inference at t=0 (or t=1) is missing. Is the performance gain purely from smoother manifold learning, or a specific property of the flow path?
2.Absence of clinical downstream evaluation. While the quantitative results (MAE, PSNR, SSIM) are strong, the evaluation would be more impactful with clinical tasks. For Task 1 (MR-to-PET), an analysis of SUVR across diagnostic groups (AD, MCI) is recommended. For Task 2 (T1w-to-T1ce), a downstream tumor segmentation task would better validate the preservation of clinically relevant features.
Minor 1: In Fig. 2(a), the validation loss of the RM baseline does not show a overfitting curve, though the advantage of JiR in sustained learning is evident. Minor 2: In Fig. 3 (Task 1), the visual differences in the regions indicated by yellow arrows are subtle. Additional explanation would help clarify the superiority of the proposed method.
- 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 submission does not provide sufficient information for reproducibility.
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper provides a foundational analysis of one-step medical image translation by identifying the time-condition t as a key deterministic regularizer. The proposed method is novel, offering a clear critique of existing frameworks while highlighting the distinct contributions of this study. Although some minor analyses are brief due to page limits, all critical components are articulated with clarity.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The core premise of this paper is to simplify the diffusion process by treating the time component as an image regularizer, referred to as JIR. The domain under investigation is medical image translation, also known as synthesis or contrast synthesis. The overall goal is to reduce hallucinations when applying diffusion-like processes for image translation. The method builds on flow networks, and the approach is validated on MR-to-PET data from the Alzheimer’s Disease Neuroimaging Initiative as well as multi-contrast MR data in rats. For memory efficiency, both datasets are cropped and resampled. The main contribution of this paper is an innovative formulation and design that improves the interpretability of synthesis networks. The method is validated on widely available and relevant datasets, and the technical details are presented at a level that supports reproducibility.
- 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.
The major strengths of the paper include a clear vision, a strong theoretical foundation, and meaningful architectural innovation. In addition, the performance is exceptional on widely recognized benchmark datasets. The manuscript is well written and accessible, while still maintaining appropriate technical depth.
- 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.
There are a few minor weaknesses to address. The brain images are not consistently presented in standard anatomical orientation, and in Figure 3, the top panel appears to be flipped upside down. It would have been better if the data could have been used as provided (not resampled/cropped). In addition, no code is currently provided, which may limit reproducibility.
- 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 submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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
Overall, this paper is both highly novel and exceptionally clear in its presentation of the theoretical framework. I am very enthusiastic about this work and would support its acceptance.
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall, this paper is both highly novel and exceptionally clear in its presentation of the theoretical framework. I am very enthusiastic about this work and would support its acceptance.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We sincerely thank the reviewers and meta-reviewer for the positive and constructive feedback. We are encouraged that the reviewers recognized the novelty, clarity, and relevance of our work. We have addressed the suggested clarifications in the final version.
For Reviewer #1, (1) ω(t) in Eq. (6) is inspired by the x-prediction formulation in [16], but adapted to medical image translation. Compared with the original design, we reduce the weight oscillation and emphasize stable reconstruction, so that training focuses more on faithful translation rather than denoising. (2) The sampling steps in Fig. 2 refer to inference-time ODE solver steps, as in standard diffusion/flow models. We use the Heun solver for multi-step sampling, while JiR is designed for deterministic one-step inference. For BiF, more sampling steps may accumulate unnecessary perturbation from the bridge path, which can slightly degrade fidelity. (3) Regarding efficiency, JiR uses NFE=1 and is therefore much faster than conventional multi-step diffusion/flow models, while remaining architecture-agnostic and nearly as efficient as regression models.
For Reviewer #2, (1) Thank you for your insightful comments. We suggest that the benefit of JiR is synergistic. Time-conditioned training supervises a continuum of intermediate states rather than only a direct endpoint mapping, which provides structured regularization. BiF and x-prediction further remove unnecessary stochasticity and align the framework with fidelity-driven medical translation, making JiR directly comparable to one-step regression while retaining the benefit of diffusion-style training. (2) We fully agree that downstream evaluations are important. Due to space and scope, we focused on methodological analysis and controlled comparisons, and have mentioned downstream clinical validation as future work. (3) For Fig. 2(a), the RM baseline may not show dramatic divergence, but its validation loss saturates much earlier than JiR, supporting the regularization effect.
For Reviewer #4: Thank you for pointing out the orientation issue in Fig. 3.We have corrected the figure and ensured consistent anatomical orientation. Cropping and resampling were used due to memory constraints in 3D volumetric training, and the same preprocessing was applied to all methods for fair comparison. We have also made our code publicly available in the final version.
Finally, following the meta-reviewer’s suggestion, we have checked and updated the references, replacing outdated or incomplete arXiv citations with peer-reviewed versions whenever available.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
All three reviewers agree upon the contribution and novelty as well as the clarity and quality of the paper. Reviewers have mentioned a few smaller issues which I invite the authors to consider for the final paper, to further improve paper quality.
I also expect that the references are cleaned up as there are many incomplete ones in the current manuscript version, and additionally there are many citations to arxiv of papers which likely have been published at a peer reviewed venue.