List of Papers Browse by Subject Areas Author List
Abstract
Surgical endoscopy inherently involves constrained lighting conditions, which can obscure fine anatomical structures. Consequently, low-light image enhancement (LLIE) is crucial for improving visual clarity and structural discernibility. Recent diffusion-based methods improve restoration quality but are typically trained from scratch on small medical datasets, limiting their ability to exploit broader visual priors. Large-scale pretrained text-to-image (T2I) diffusion models encode powerful generative knowledge, yet directly applying them to surgical LLIE suffers from mismatched visual characteristics and high inference cost. We propose S2DLight, a single-step latent restoration framework built on a Stable Diffusion backbone for endoscopic LLIE. We introduce Degradation-Aware Adaptation to provide image-specific low-light guidance via contrastive degradation representations, and reformulate iterative sampling into one forward latent prediction to improve efficiency. To compensate for removing progressive refinement, we incorporate Latent Structural Augmentation, and further mitigate training-inference prompt mismatch using Training-Time Semantic Alignment. Experiments on public endoscopic benchmarks demonstrate that S2DLight achieves state-of-the-art performance while reducing the computational burden associated with large-scale diffusion inference.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2353_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/FeiSong123/S2DLight
Link to the Dataset(s)
N/A
BibTex
@InProceedings{FeiSon_S2DLight_MICCAI2026,
author = { Fei, Song AND Wang, Guanyi AND Ye, Tian AND Chen, Sixiang AND Lai, Jianyu AND Zhu, Lei},
title = { { S2DLight: Single Step Diffusion Network via Degradation-Aware Adaptation for Surgical Endoscopic Image Low-Light Enhancement } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16892},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
S2DLight introduces a single-step diffusion framework specifically tailored for surgical endoscopic image enhancement. By integrating Stable Diffusion pre-trained priors into the medical imaging domain, it systematically addresses the bottlenecks of conventional diffusion models in surgical contexts, such as slow inference, mismatched visual features, and the absence of semantic guidance.
- 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 core strength of S2DLight lies in its systematic adaptation of large-scale pre-trained T2I diffusion models (Stable Diffusion) for surgical endoscopic enhancement. It ingeniously addresses the clinical bottleneck of slow inference through single-step reformulation, while integrating degradation-aware adaptation and semantic alignment to bridge the domain gap between medical and natural imagery. This framework provides a practical and efficient solution for deploying diffusion models within real-time surgical workflows.
- 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 manuscript does not explicitly explain the rationale for why using a trained degradation estimation network is more effective at capturing low-light characteristics. 2.Regarding the training of the Degradation Estimation Network, the paper utilizes two patches from the same low-light image as positive pairs and patches from other images as negative samples for contrastive training. However, there is no explanation justifying this specific approach. What is the underlying rationale for this contrastive learning method? What specific problem is it intended to address, and how or why does this training enable the model to resolve such issues? 3.The single-step diffusion mentioned in the paper appears to resemble the DDIM (Denoising Diffusion Implicit Models) algorithm. Could the authors clarify the distinctions between the proposed method and DDIM, as well as its specific advantages?
- 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 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.
(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?
While the paper demonstrates a substantial amount of work, several modules are insufficiently explained.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 modules still lack interpretability and innovation.
Review #2
- Please describe the contribution of the paper
The authors propose S2DLight, a low-light image enhancement (LLIE) method for restoring surgical endoscopic images that reuses a pre-trained text-to-image Stable Diffusion model, introducing intermediate adaptations and LoRA-based tuning. The experiments show clear quantitative improvements over the state of the art (SOTA) while maintaining competitive inference times for near-real-time applications.
- 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 paper is very well written, featuring a clearly explained methodology, excellent visualizations of the processing pipeline, and thorough qualitative evaluations and comparisons. The method addresses a real-world problem with well-posed motivation, the ablation study effectively justifies the pipeline’s design choices, and the authors commendably retrain a non-surgical LLIE method to ensure a fair comparison.
- 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.
While the quantitative results show a significant improvement over SOTA, the qualitative evaluation doesn’t really show differences or improvements over existing methods, such as LighTDiff or DiffLL, raising the question whether the additional computational costs and longer inference time are justified for the visual quality improvements. Reusing a pre-trained diffusion model but adapting it for single-pass inference raises the question of why a simpler, inherently non-diffusion-based approach was not used instead. Additionally, while the ablation study shows that Latent Structural Augmentation successfully bypasses the need for iterative refinement and improves performance, the paper does not demonstrate the performance difference between using standard diffusion refinement versus relying solely on this augmentation.
- 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
1.What is the primary motivation for using a diffusion model if inference is ultimately performed in a single iteration? 2.Since generating training data via [6] alters colors, how does the model perform in real-world settings where colors are not distorted by gamma transformations? 3.Finally, how would the model perform if the degradation embedding branch (DENet) were removed or replaced with a fixed embedding?
- 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?
While the method is well-motivated and demonstrates quantitative improvements, the lack of noticeable qualitative enhancements over existing methods does not currently justify the increased computational cost of a diffusion-based approach, making the overall benefit marginal.
- 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 #3
- Please describe the contribution of the paper
This work’s main contribution is S2DLight, a single-step diffusion framework that adapts a pretrained text-to-image diffusion model to surgical endoscopic low-light image enhancement. With Degradation-Aware Adaptation, Single-Step Reformulation, and Training-Time Semantic Alignment, it improves enhancement quality while significantly cutting inference cost.
- 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.
This paper addresses surgical endoscopic low-light image enhancement, a task with clear clinical relevance, by proposing S2DLight, a single-step diffusion framework. The overall framework is well designed and clearly problem-driven, and it effectively tackles several key challenges in adapting pretrained text-to-image (T2I) diffusion models to the medical domain, including domain discrepancy, inference efficiency, and semantic-conditioning inconsistency. The comparative experiments and ablation studies on public datasets are sufficiently thorough and provide solid evidence for the effectiveness of the proposed method. Particularly commendable is the Single-Step Reformulation, which significantly reduces inference cost while largely preserving enhancement quality, making the method not only technically interesting but also promising for practical 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.
Although the paper reports strong quantitative results on the relevant datasets, it still has several limitations. First, the method is evaluated only on EndoVis17/18 and has not been validated in real clinical low-light scenarios. Second, no downstream tasks are included, which makes the clinical applicability of the method less convincing.
- 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 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?
Overall, this work is not the kind of very solid work that would warrant a strong accept, but it is clearly better than a weak reject. It falls into the category of having methodological highlights and sufficient experiments, yet lacking external validation and application support. Therefore, I give it a weak accept.
- 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.
I would recommend accepting this paper as a poster. The lack of real clinical low-light data and downstream task validation is still a limitation, but I do not think it undermines the main technical contribution or the experimental results presented in the paper. Given the strong quantitative performance, reasonably complete ablation studies, and the overall reviewer scores being slightly above the acceptance bar, I think the paper meets the standard for acceptance.
Author Feedback
We sincerely thank all reviewers for their constructive and insightful comments. We appreciate their recognition of the relevance, motivation, and experimental quality of S2DLight. The comments are valuable for improving the clarity of our manuscript. For clarity, W denotes weakness items and C denotes optional comments.
[Common questions]
[R1-W3, R2-W2/C1] Motivation for diffusion and difference from DDIM R: Our goal is not iterative sampling, but transferring pretrained Stable Diffusion priors to surgical LLIE, where medical data are limited. S2DLight differs from DDIM. DDIM follows a scheduled reverse trajectory. In contrast, S2DLight treats the encoded low-light latent as a degraded state at a fixed timestep and restores it in one forward pass. It neither starts from Gaussian noise nor recursively denoises the latent. Thus, S2DLight is a single-step latent restoration framework using SD priors, not a DDIM sampler.
[R1-W1/W2, R2-C3] DENet and contrastive degradation learning R: DENet provides image-specific degradation guidance. Low-light degradation varies across frames due to illumination, exposure, noise, and tissue appearance; thus a fixed embedding cannot adapt to different degradation severity. In contrastive learning, patches from the same low-light image are positives because they share similar degradation, while patches from different images are negatives because their degradation statistics usually differ. This encourages degradation-related cues and suppresses local anatomy-specific information. Removing DENet or using fixed embedding would reduce the model to a generic prompt-conditioned enhancer.
[R2-C2, R3-W1/W2] Synthetic degradation, clinical validation, and downstream tasks R: We agree that real clinical low-light data and downstream evaluation would further strengthen the evidence. However, paired real low-light/normal-light surgical data are difficult to obtain, since collecting well-aligned pairs under different illumination settings in real procedures is challenging. Thus, synthetic paired protocols are commonly used in surgical LLIE for training and fair comparison. We will state this limitation, avoid over-claiming clinical deployment, and discuss real-data collection, unpaired/semi-supervised adaptation, and downstream tasks as future work.
[Reviewer-specific questions]
[R1] Modules insufficiently explained R: We appreciate this comment and will better explain the modules: DENet provides image-specific degradation conditioning; contrastive learning separates degradation cues from anatomy; Single-Step Reformulation transfers SD priors without iterative sampling; and Latent Structural Augmentation compensates for reduced progressive refinement.
[R2-W3] Standard diffusion refinement vs proposed single-step design R: The submitted manuscript includes 1/4/50-step comparisons, directly evaluating iterative refinement against our single-step design. Results show that S2DLight achieves better overall accuracy with much lower inference cost. This suggests that directly restoring the encoded low-light latent is more effective for surgical LLIE than repeated diffusion refinement, which may cause over-refinement or distribution drift.
[R2-W1] Limited qualitative improvement R: We understand this concern. Differences between strong LLIE methods can be hard to observe due to limited figure size and fine endoscopic structures. We will add enlarged local regions. The consistent gains in PSNR, SSIM, and LPIPS further support improved fidelity, structure preservation, and perceptual quality.
[R3-W2] Lack of downstream task evaluation R: We agree that downstream tasks such as segmentation or instrument recognition would provide additional clinical evidence. This work focuses on surgical LLIE, and our experiments evaluate paired enhancement quality, efficiency, and component effectiveness. We will avoid over-claiming clinical deployment and include downstream validation as future work.
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.
This paper proposes a single step Stable Diffusion based framework for low light enhancement of surgical endoscopic images, using degradation aware conditioning, latent structural augmentation, and training time semantic alignment to improve quality and speed. The reviewers noted the promising performance but raised several important limitations. First, the methodological motivation remains unclear, particularly the rationale for the degradation estimation network, the contrastive learning design, and the distinction between the proposed single step formulation and existing DDIM style diffusion methods. Second, there is a disconnect between the strong quantitative gains and the qualitative examples, making it unclear whether the improvements translate into visibly meaningful enhancement or justify the computational complexity of a large diffusion backbone. Third, the experimental validation is limited to synthetic low light degradation on EndoVis17 and EndoVis18, without evaluation on real clinical low light scenarios or downstream surgical tasks, which weakens the practical and clinical claims. Decision is rebuttal.
- 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.
After rebuttal, I recommend accept decision. The reviewers agree that the paper addresses a relevant surgical image enhancement problem with strong quantitative results, fair comparisons, and useful ablation studies. The main concerns relate to novelty, interpretability of individual modules, qualitative improvement over existing LLIE methods, and limited clinical/downstream validation. The rebuttal provides reasonable clarification of the distinction from DDIM, the motivation for transferring Stable Diffusion priors rather than using iterative sampling, and the role of DENet and contrastive degradation learning. While the lack of real clinical low-light validation and downstream task evaluation remains as limitations, these issues are adequately acknowledged and do not outweigh the methodological contribution and experimental evidence. As a result, the decision is accept.
Meta-review #2
- 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.
Orignally 4-4-3.where the 3 was the only reviewer with post-rebuttal interaction, and remained the reject. Having carefully looked at the reviews, the rebuttal and the paper, I feel that this is the most borderline case I have seen so far. Reviewers rightfully point out the lack of thorough clinical evaluation, but also see the technical merit. I would give this paper the benefit of the doubt.
Meta-review #3
- 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.
After the rebuttal, R1 maintained a Reject recommendation, R3 maintained an Accept recommendation, while R2, who initially provided a Weak Accept recommendation, did not provide an updated review. Although the paper demonstrates some strengths, including promising experimental results, it also exhibits several fundamental limitations.
First, as noted by both R1 and R3, the proposed model was evaluated only on synthetic low-light data, leaving its real-world clinical utility unclear. While collecting paired low-light and normal-light clinical data may be challenging, it would still be feasible to collect low-light clinical data and provide qualitative or quantitative results to better assess the model’s practical performance.
Second, the rebuttal does not fully address R2’s concern regarding the necessity of the diffusion-based design. As stated in the rebuttal, the proposed framework neither starts from Gaussian noise nor recursively denoises latent representations, but instead operates as a single-step latent restoration framework using Stable Diffusion priors. This raises the question of whether a simpler denoising autoencoder, initialized with pretrained Stable Diffusion components, could achieve similar performance with a more straightforward design and justification.
Considering these concerns, the paper is not recommended for acceptance. The authors are encouraged to address the reviewers’ comments when revising the paper for submission elsewhere.