- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes RePriFusion, a low-light image enhancement framework that combines Retinex-based illumination and reflectance features with pretrained DINOv3 representations. The reported results demonstrate competitive restoration quality and promising cross-dataset generalization.
Critical strengths
- Effective integration of complementary visual priors.
- Strong empirical performance.
- Reasonably comprehensive ablation studies.
Critical weaknesses
-
The paper emphasizes explicit Retinex reconstruction as a central contribution. However, according to Eqs. (5)–(8). Since neither illumination nor reflectance is explicitly constrained by a decomposition loss or physical regularization, their intended interpretations are not guaranteed. The reported ablation demonstrates the usefulness of the reconstruction pathway, but does not establish that the improvement comes from physically meaningful Retinex decomposition rather than additional nonlinear processing and residual correction.
-
The methodological novelty of prior-specific fusion remains somewhat limited. Although Tables 4 and 7 support the selected design, the stage configurations are not strictly capacity-matched, and several differences between fusion alternatives are relatively small. It remains unclear whether the improvement comes from the proposed prior-specific design principle or simply from additional pretrained features and carefully selected architectural configurations. A stronger comparison with a parameter-matched generic multi-prior fusion baseline would better establish the contribution.
-
The model is trained jointly on five paired datasets with noise augmentation, but only a few competitors are retrained under this same protocol. It is unclear whether the reported gains stem from architectural novelty or simply from a richer training distribution. To prove the actual architectural merit, an apples-to-apples comparison is needed. Please retrain strong, efficient baselines (e.g., CIDNet, Multinex) under the exact same joint-training and augmentation setup.
-
The framework is quite heavy, stacking a Retinex estimator, multi-stage fusion, and a frozen DINOv3 encoder. Even the "Small" variant hits ~287 GFLOPs, and the Base model jumps to ~810 GFLOPs—substantially higher than many competitive methods. While restoration metrics improve, it is unclear if these marginal gains justify the massive inference overhead. Given the relatively small differences between some fusion configurations, these results require repeated-run validation. The authors must provide a clear performance-efficiency trade-off analysis, including actual inference latency and memory footprint, to prove this framework is worth the cost.
Questions for authors
How can the authors demonstrate that the learned illumination and reflectance components represent meaningful Retinex priors rather than unconstrained intermediate features?
Can the authors compare prior-specific fusion against a parameter-matched generic fusion strategy to better isolate its contribution?
How does RePriFusion compare with representative efficient baselines when all methods use the same mixed-dataset training and augmentation protocol?
What are the measured inference latency and memory requirements, and how stable are the reported improvements across independent training runs?
The new three-seed, capacity-matched results address much of my concern about the benefits of prior-specific fusion. The CRPE comparison still changes both the inputs and the residual parameterization, making it difficult to isolate the contribution of the Retinex formulation. I retain my rating of 2 (Weak rejection).