acceptodds
Under review as a conference paper at ICLR 2027

UG-Prior: Uncertainty-Guided Priors with Online Refinement for Zero-Shot Low-Light Image Enhancement

Abstract

Pixel-level and distribution-level supervision for low-light image enhancement (LLIE) remain sensitive to scene and degradation in the training data, limiting generalization. Physical illumination-invariant priors bridge low/normal-light domains, posing a promising direction for improved generalization. However, their unreliable estimation from degraded observations hampers cross-domain robustness and faithful enhancement. To alleviate this issue, we propose UG-Prior, an Uncertainty-Guided zero-shot LLIE framework. Specifically, we propose uncertainty-guided multi-scale fusion to construct robust and informative priors with reliable structural and chromatic cues. An auxiliary decoder is proposed to regularize generative features by reconstructing clean intrinsic representations. During inference, its predictions enable online prior-image co-optimization, which refines the conditioning prior as the image evolves. Extensive experiments demonstrate that our method achieves state-of-the-art zero-shot LLIE. Relative to the recent SOTA, ZeroIDIR, our UG-Prior improves PSNR by 0.42 dB and LPIPS by 0.02 on LOL, with strong generalization across diverse scenes and illumination conditions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
zero-shot low-light enhancement, illumination-invariants, generalization

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_mrSqOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This article proposes UG Prior, an uncertainty guided framework for zero sample low light image enhancement. The author points out that the existing illumination invariant priors may have unreliable estimates in severely degraded low light environments, making it difficult to balance robustness and information richness. To address this issue, the author proposes uncertainty guided multi-scale prior fusion in this article and introduces a color reliability weighting mechanism to preserve reliable structural and color information. The author designed a prior decoding regularization and online prior image collaborative optimization mechanism to further improve the prior utilization and dynamic updating in the generation process. The experiment was conducted on multiple low light image enhancement datasets, and the results showed that the method has good performance in image restoration quality and cross dataset generalization.

Critical strengths

The research method design is targeted. This article focuses on the reliability of illumination invariant priors under severe low illumination degradation conditions, and balances the robustness and information richness of priors through uncertainty guided multi-scale fusion. Compared to directly using a fixed lighting invariant prior, this adaptive construction method has a reasonable motivation and also reflects a certain degree of innovation.

The method design is complete and the experimental results are convincing. This article not only improves prior construction, but also enhances prior utilization in the generation process through prior decoding regularization and online update mechanisms. The author conducted quantitative and qualitative comparisons on multiple datasets and provided ablation experiments, visualization analysis, and model distillation experiments. The overall results support the effectiveness and certain cross dataset generalization ability of the proposed method.

Critical weaknesses

The effectiveness of uncertainty estimation still lacks direct verification. This article constructs uncertainty maps through brightness and local structure statistics, and uses them to guide multi-scale prior fusion. Although ablation experiments have shown that this design can improve the final enhancement performance, there is still a lack of direct evidence to prove a good correspondence between the estimated uncertainty and the true prior error. This to some extent weakens the explanation of the core mechanism.

The effectiveness of the online prior image collaborative optimization mechanism still requires more direct evidence. This article predicts priors through an auxiliary decoder and performs weighted updates during the generation process. Although ablation experiments have shown that online updates can improve the final results, it has not been fully demonstrated whether prior representations gradually become more accurate with iteration. In addition, this update mechanism is mainly based on weighted fusion, and its difference from general iterative prior correction methods can be further clarified.

Questions for authors
  1. Section 3.2 estimates uncertainty using brightness and structural information. Has the correlation performance between the uncertainty graph and the true prior error been verified?

  2. Section 3.4 achieves online prior updates through weighted fusion. Can the prior error changes under different generation steps be displayed to verify whether the updated prior is indeed more accurate?

  3. When using QuadPrior prior in Table 3, the PSNR decreased from 21.29 dB to 19.10 dB. Did the experiment maintain the same generation backbone and training settings to ensure that the performance difference mainly came from the prior itself?

  4. Appendix D.2 mentions adjusting the output of other methods uniformly to 1024 resolution. Will this scaling affect the evaluation results when using the same test images and partitions for all methods?

  5. Defogging and underwater image enhancement can use the same theory as low light image enhancement. Can your paper verify the effectiveness of these two tasks. Note: SOTA is not required here, only the universality of your framework needs to be verified.

I suggest accepting this article clearly. But you need to discuss the above issues in detail.

6 comments
AuthorsOpenReview ↗

Official Response to Reviewer mrSq (Part 1/3)

We appreciate the reviewer’s positive assessment and constructive suggestions for directly validating uncertainty–error correspondence, the accuracy of online prior refinement, and broader generalization capabilties on dehazing/under water enhancement.

Q1 / Weakness 1: Direct evidence to verify the correspondence between uncertainty and prior error.

We provide additional analysis demonstrating the correspondence between estimated uncertainty and empirical prior error. As described in our response to R1-Q1, we evaluate 291 short–long exposure pairs from 49 SID Fuji scenes [2], with three short exposures (33/40/100 ms), five Gaussian scales, and three patch sizes (8/16/32). At each scale, prior error is computed as the MSE between the patch-normalized short- and long-exposure priors. The matched long exposure provides the empirical reference prior. Uncertainty is computed from the short exposure images alone.

Because uncertainty is u=1−qu=1-q, its correlation with error has the opposite sign to the confidence–error correlation. All 45 scene-averaged exposure–scale–patch-size correlations are positive. Representative Pearson coefficients are shown below. Each cell reports the fine/coarse scales, σ=0.5/2.5\sigma=0.5/2.5.

Patch size33 ms40 ms100 ms
8 × 80.319 / 0.1350.257 / 0.0580.459 / 0.089
16 × 160.446 / 0.2010.349 / 0.0800.582 / 0.116
32 × 320.499 / 0.2590.390 / 0.0950.612 / 0.138

These results directly support uncertainty as an empirical reliability cue, with a stronger association at fine scales. Therefore, downweighting fine-scale responses in unreliable regions while retaining their detail in reliable regions intuitively improves the robustness–informativeness trade-off, as shown in Figure 3. We provide a more straightforward figure in Appendix A.1.


Q2 / Weaknesses 2: More direct evidence for prior refinement mechanism.

To explore whether prior representations gradually become more accurate with iteration, we record the evolving prior at every flow timestep on 15 LOL-v1 test images over 28 sampling steps [3]. Its error is measured against the normal-light reference prior and divided by the initial prior error for each image.

The table reports mean relative prior MSE ↓. Note that the relative error remains 1 during steps 1–8 because the first updated prior is used at step 9. The final optimized prior has 33.8% lower error than the initial prior, with improvement on all 15 images. This experiment provides further evidence for the effectiveness of the proposed online prior–image co-optimization mechanism.

Sampling step1–89122028
Frozen prior1.0001.0001.0001.0001.000
Online refinement1.0000.6850.6720.6670.662

Compared to general iterative prior correction methods, our methodological distinction lies in the learned interaction between image generation and prior refinement, rather than the fusion formula itself. Our mechanism refines the physical conditioning prior through bidirectional interaction with the evolving generative features with all network weights fixed and no test-time backpropagation. This differs from test-time optimization in existing generative restoration methods. We have included this experiment in Sec.4.5, Page 10.


Q3: Were the backbone and training settings controlled in Table 3?

Yes. The 21.29-versus-19.10 dB comparison uses the same FLUX generation backbone, training data, and training hyperparameters, exclusively replacing the conditioning prior with QuadPrior's [1]. We have clarified this in the main text. This result reveals a counterintuitive phenomenon: a stronger backbone does not necessarily compensate for a less suitable condition. One possible explanation is that a high-capacity generator learns perturbation-specific mappings when the prior inadequately balances robustness and information retention. In contrast, our prior enables more effective conditioning under the controlled setting.

AuthorsOpenReview ↗

Official Response to Reviewer mrSq (Part 2/3)

Q4: Does resizing affect the evaluation?

We apologize for the incorrect description of the evaluation resolution. All comparison methods use their official inference commands, and outputs are resized to the original GT dimensions for evaluation. UG-Prior generates images at 1024 × 1024, followed by bicubic resizing using PIL.Image. We use default pyiqa=0.1.15.post2 to calculate metrics.

To ablate the effect of resizing, we evaluate the same 28-step predictions on all 115 LOL images under two protocols: (A) native 1024 × 1024 outputs against bicubic-upsampled GT, (B) outputs bicubic-resized to the original GT dimensions against unchanged GT.

Compared to resized evaluation, our original resolution evaluation leads to PSNR and SSIM generally decreases but LPIPS [7] all improves. These differences reflect resampling and evaluation-scale effects. We have clarified this protocol in the appendix.

MetricA: 1024 × 1024B: Original GT dimensionsΔ (B-A)Images improved
RGB PSNR ↑ (dB)21.3421.29-0.05045/115
SSIM-YIQ ↑0.8580.838-0.02086/115
LPIPS-AlexNet ↓0.2300.159−0.071115/115

Q5: Generalization to dehazing and underwater enhancement.

Thank you for this constructive suggestion. We have added quantitative and qualitative results in Appendix A.9, and copy the main results here. These tasks share the challenge of recovering scene content under degraded visibility, while scattering and wavelength-dependent attenuation require task-specific adaptation.

Evaluation setting. We evaluate the our UG-Prior model on fixed exploratory subsets of 16 O-HAZE images and 30 dark UIEB images, without task-specific fine-tuning.

Dehazing. We equipped UG-Prior with WB and DCP as preprocessor to adapt the frozen low-light enhancement model to accept hazy inputs. We evaluate our method and three task-specific methods (DehazeFormer, DEA-Net, DNMGDT) with three full-reference metrics and four non-reference metrics. The comparisons use the official outdoor DehazeFormer-B and DEA-Net checkpoints and the released DNMGDT checkpoint. The qualitative and quantitative comparison are shown below:

MethodNIQE ↓BRISQUE ↓MUSIQ ↑CLIP-IQA ↑PSNR ↑ (dB)SSIM ↑LPIPS ↓
DehazeFormer, TIP'232.1878.30766.5150.368217.3260.70580.3057
DEA-Net, TIP'242.1979.16266.8310.393517.1990.71850.2958
DNMGDT, TMM'254.86034.52253.8160.350818.3330.70700.4291
Ours2.19410.34564.4330.451218.0030.75390.2875

We provide additional visualized results and limitation analysis in Appendix A.9 Fig.20 and Fig.21.

Underwater enhancement. We evaluate the frozen UG-Prior model without underwater-specific fine-tuning on a fixed 30-image UIEB subset. We compare against three underwater-trained models including QualiTeacher (AIM-Net-based) [ECCV'26] [11] PUIE-Net [ECCV'22] [12] and NU2Net [AAAI 2023] [13] using its official pretrained underwater checkpoint and inference procedure.

MethodLPIPS ↓NIQE ↓BRISQUE ↓MUSIQ ↑CLIP-IQA ↑
PUIE-Net, ECCV'220.17024.350730.592243.03760.2827
NU2Net, AAAI'230.26794.792338.913136.72450.2733
QualiTeacher, ECCV'260.25514.241025.132456.03110.3960
UG-Prior0.30113.460613.346455.51350.4346

UG-Prior achieves the best on three of the four no-reference metrics and the second best on the remaining one. Specifically, it obtains better per-image scores on 26/30 images for NIQE, 26/30 for BRISQUE, 15/30 for MUSIQ, and 21/30 for CLIP-IQA. Although visually plausible, as UG-Prior use no paired UIE data during training, it is hard to establish superior restoration fidelity required by the exact ground-turth images. We provide visualized comparison in Appendix A.9 Fig.22 and discuss more, verifying the universality of our method.

AuthorsOpenReview ↗

Official Response to Reviewer mrSq (Part 3/3)

References

  1. Wenjing Wang et al. Zero-Reference Low-Light Enhancement via Physical Quadruple Priors. CVPR, 2024.
  2. Chen Chen et al. Learning to See in the Dark. CVPR, 2018. SID dataset.
  3. Chen Wei et al. Deep Retinex Decomposition for Low-Light Enhancement. BMVC, 2018. LOL dataset.
  4. Ben Fei et al. Generative Diffusion Prior for Unified Image Restoration and Enhancement. CVPR, 2023.
  5. Huaqiu Li et al. LD-RPS: Zero-Shot Unified Image Restoration via Latent Diffusion Recurrent Posterior Sampling. ICCV, 2025.
  6. Zhou Wang et al. Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE Transactions on Image Processing, 2004.
  7. Richard Zhang et al. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. CVPR, 2018.
  8. Codruta O. Ancuti et al. O-HAZE: A Dehazing Benchmark With Real Hazy and Haze-Free Outdoor Images. CVPR Workshops, 2018.
  9. Chongyi Li et al. An Underwater Image Enhancement Benchmark Dataset and Beyond. IEEE Transactions on Image Processing, 2019.
  10. Kaiming He, Jian Sun, and Xiaoou Tang. Single Image Haze Removal Using Dark Channel Prior. CVPR, 2009.
  11. Fengyang Xiao et al. QualiTeacher: Quality-Conditioned Pseudo-Labeling for Real-World Image Restoration. ECCV, 2026.
  12. Zhenqi Fu et al. Uncertainty Inspired Underwater Image Enhancement. ECCV, 2022.
  13. Chunle Guo et al. Underwater Ranker: Learn Which Is Better and How to Be Better. AAAI, 2023.
  14. Yuda Song et al. Vision Transformers for Single Image Dehazing. IEEE TIP, 2023.
  15. Zixuan Chen et al. DEA-Net: Single Image Dehazing Based on Detail-Enhanced Convolution and Content-Guided Attention. IEEE TIP, 2024.
  16. Yanzhao Su et al. Real Scene Single Image Dehazing Network With Multi-Prior Guidance and Domain Transfer. IEEE TMM, 2025.
Reviewer_83b1OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates zero-shot low-light image enhancement. Addressing the robustness–informativeness trade-off where analytical, physics-based illumination-invariant priors become noisy and unreliable under severe real-world low-light degradation, the authors present UG-Prior.


Gaussian scale-space representations are adaptively fused using spatial confidence and uncertainty derived from local luminance and structure-tensor eigenvalues, effectively discriminating directional coherence from isotropic gradient noise. Bilateral filtering and luminance-dependent reliability gating are applied to chromatic cues in HSV space.


Leveraging FLUX.1-dev as a generative backbone, an auxiliary U-Net decoder reconstructs clean intrinsic targets from intermediate generative tokens during training on synthetic perturbations of normal-light images, preventing degradation-sensitive features from dominating the condition stream.


During Flow Matching sampling at inference time, feedback from the evolving generative state is iteratively incorporated into the intrinsic prior in closed form via the auxiliary decoding proxy, bypassing costly per-step pixel decoding and re-encoding.


The expensive generative teacher model is distilled into a lightweight Restormer-based student architecture equipped with Spatial Feature Transform layers for practical deployment.

Critical strengths

The appendix provides a complete derivation of multi-scale Gaussian color invariants under idealized linear sensor assumptions, while honestly articulating theoretical boundaries introduced by numerical offsets, spatial non-uniformity, and additive perturbations.


The formulation of prior–image co-optimization as a regularized least-squares problem with a closed-form step update enables dynamic condition refinement during diffusion sampling without the prohibitive overhead of full image synthesis at each step.


UG-Prior establishes state-of-the-art performance among zero-shot and unsupervised baselines across LOL (21.29 dB PSNR vs. QuadPrior's 20.31 dB and ZeroIDIR's 20.87 dB), LSRW, FiveK, BAID, and unpaired benchmarks. The distilled student network achieves a substantial FLOP reduction and drops peak memory from 32 GB to 0.538 GB with minimal metric penalty.

Critical weaknesses

The teacher framework relies on a heavy FLUX.1-dev foundation model. On a modern NVIDIA H20 GPU, processing a single 1024×10241024 \times 1024 image requires 53.5 seconds, which severely restricts direct end-to-end applicability in real-world pipelines prior to distillation.


When low-light observations suffer from near-zero photon counts (clipping at pixel value 0), physical information is completely lost, causing generative priors to fabricate hallucinated structural artifacts, as acknowledged in the failure analysis.


Training relies entirely on synthetic noise perturbations (Gamma, Gaussian, and Poisson) applied to normal-light images. While this avoids real low-light training pairs, it remains an open question how well the model handles intricate sensor-specific read noise and non-stationary ISP artifacts.

Questions for authors

Given that the teacher model requires 53.5 seconds per 1024×10241024 \times 1024 image on an NVIDIA H20 GPU, how critical is the massive model capacity of FLUX.1-dev to the quality of the pseudo-targets during distillation? Could a significantly more lightweight generative foundation model (e.g., SD-Turbo, SDXL-Lightning, or a compact DiT) serve as the teacher without degrading the performance of the distilled student network?


The training perturbation pipeline relies entirely on synthetic Gamma adjustments, Gaussian noise, and Poisson noise applied to normal-light images. Have the authors evaluated the teacher or student model on real-world raw-to-sRGB datasets (such as SID or ELD), where noise exhibits intricate sensor-specific read/shot noise characteristics and non-linear ISP processing?


In Equations 11 and 12, the online refinement schedule fixes the threshold τ=0.25\tau = 0.25 and the step weight α=0.2\alpha = 0.2. How sensitive is the final visual quality to these heuristic hyperparameter values across drastically different lighting conditions? Could α\alpha be dynamically adjusted per image based on the estimated spatial uncertainty map uu?


For regions suffering from near-zero photon counts where physical cues are completely lost (e.g., pixel value = 0 in Fig. 9), does the framework provide any confidence metric or uncertainty mask to warn downstream safety-critical applications (such as autonomous night driving or surveillance) of potentially hallucinated content?


6 comments
AuthorsOpenReview ↗

Official Response to Reviewer 83b1 (Part 1/3)

We thank the reviewer for raising concerns about computational cost and real-world reliability, particularly whether a lighter teacher can preserve student performance.

Q1 / Weakness 1: Teacher cost comparison and lightweight distilled models.

We agree that the teacher's runtime of 53.5 s/image on an H20 limits direct deployment. We therefore evaluate 1–28 sampling steps to characterize its quality–latency trade-off and report distilled student performance under matched settings to assess practical utility.

1. Quality–latency trade-off across sampling steps

We evaluate sampling budgets of 1–28 steps on the same 115 LOL test images [1] and present the quality–latency curves in Fig. 4 of the revised manuscript. These runtimes are measured on NVIDIA GeForce RTX 4090-48G.

The results show a flexible quality–latency trade-off:

  • Notably, with only one sampling step, our method achieves 20.10 dB PSNR, 0.808 SSIM, and 0.178 LPIPS in 2.63 s/image, a 15.7× speedup over the 28-step setting on the same hardware, showcasing superior performance.

  • With 4 sampling steps (6.14 s/image), UG-Prior achieves 0.169 LPIPS, improving LPIPS by 6.1% and latency by 14.5% relative to MR. Illuminate (0.180 LPIPS, 7.18 s/image).

  • With a larger budget, the best observed PSNR is 21.29 dB at 27 steps (39.90 s), which is 0.23 dB above ZeroIDIR, while the best LPIPS is 0.1565 at 17 steps (25.48 s), a 12.6% improvement relative to ZeroIDIR.

These results demonstrate that the sampling budget can be adjusted to balance enhancement quality and latency, supporting the teacher's practical utility. Additional ablation visualizations are provided in Appendix A.

2. Distilled performance

Following QuadPrior [2] and QuadPrior++ [3], we evaluate feed-forward students trained on the same 8,000 real low-light images with teacher-generated targets. With the same student architectures and training protocols, UG-Prior students achieve higher PSNR and SSIM, supporting the quality of the teacher-generated targets:

ModelLOL: PSNR ↑ / SSIM ↑LSRW: PSNR ↑ / SSIM ↑FLOPs ↓Peak memory (GB) ↓
QuadPrior teacher20.31 / 0.80816.96 / 0.5641.8 × 10¹⁴30.64
UG-Prior teacher21.29 / 0.83818.51 / 0.5754.6 × 10¹⁵32.03
QuadPrior student w/o prior18.62 / 0.76617.39 / 0.5217.750 × 10¹⁰0.385
QuadPrior student w/ SFT prior19.03 / 0.79218.28 / 0.5611.217 × 10¹¹0.412
UG-Prior student w/o prior18.81 / 0.78818.01 / 0.5457.750 × 10¹⁰0.385
UG-Prior student w/ SFT prior19.47 / 0.80318.31 / 0.5731.403 × 10¹¹0.412

Bold marks the best quality scores within the teacher and student groups separately. The UG-Prior SFT student reduces FLOPs by a factor of 3.28 × 10⁴ and peak memory by 98.7% relative to its teacher. Under the same multi-dimensional distillation framework proposed by QuadPrior++ [3], UG-Prior reaches 21.76/0.852 on LOL and 18.83/0.559 on LSRW, versus 21.75/0.825 and 18.82/0.555 for QuadPrior++.


Q2: Can a lighter teacher preserve distilled student performance?

Thank you for this suggestion. We replace FLUX.1-dev with a smaller LightningDiT [5] backbone and evaluate both the resulting teacher and its distilled student. Both teachers are trained on the same normal-light subset. For distillation, we use the same 8,000 real low-light inputs, student architecture with SFT prior injection, loss functions, and training schedule, while the supervision is generated by the corresponding teacher.

We evaluate all models on the same 115 LOL test images at the original ground-truth resolution:

ModelParameters ↓Latency ↓ (s/image)Peak memory ↓ (GB)PSNR ↑ (dB)SSIM ↑LPIPS ↓
FLUX.1-dev based Teacher (Ours)12 B41.3532.021.290.8380.159
LightningDiT-XL675 M13.416.820.500.8120.207
FLUX.1-dev based distilled student (Ours)327.36 K0.020.41219.470.8030.212
LightningDiT-XL based distilled student327.36 K0.020.41219.390.7920.220

Based on these results, changing the teacher backbone has little effect on the distilled student’s performance, suggesting that the proposed prior provides stable intrinsic conditioning across different generative backbones.

AuthorsOpenReview ↗

Official Response to Reviewer 83b1 (Part 2/3)

Q3 / Weakness 3: Real sensor noise evaluation on SID.

To add evaluation under real sensor noise, we additionally evaluate UG-Prior on 20 SID scenes at 33/40/100 ms exposures, using matched long-exposure references and no sensor-specific fine-tuning. Compared with QuadPrior, UG-Prior improves average PSNR by 0.03 dB, SSIM by 0.011, and LPIPS by 10.19%.

Although synthetic Gamma, Gaussian, and Poisson perturbations do not fully capture sensor-specific noise and ISP artifacts, these results support transfer beyond the synthetic training conditions. Together with the confidence–error analysis (more details in R1-Q1 and R3-Q1), they are consistent with our motivation: a reliable cross-illumination prior can provide useful scene guidance without explicitly modeling every degradation process.


Q4-1: Hyper-parameter sensitivity of online prior refinement across drastically different lighting conditions.

We thank the reviewer for this constructive and insightful suggestion. The scale-and-schedule ablation on 115 LOL images in Appendix A has shown that stronger updates or a longer refinement interval can reduce quality. To examine sensitivity across lighting conditions, we evaluate α ∈ {0.1, 0.2, 0.4} and τ ∈ {0.125, 0.25, 0.5} , together with a frozen-prior baseline (α = 0), on 20 SID Fuji scenes at 33/40/100 ms exposures [4]. Each cell reports Δ PSNR / Δ SSIM, where Δ = tested setting − default (α = 0.2, τ = 0.25), averages over the same 20 scenes. Positive values indicate improvement.

ατ33 ms: Δ PSNR / Δ SSIM40 ms: Δ PSNR / Δ SSIM100 ms: Δ PSNR / Δ SSIM
0 (frozen)—−0.0014 / −0.00038−0.0036 / −0.00006−0.0057 / −0.00024
0.10.125−0.0025 / −0.00016−0.0026 / −0.00019−0.0002 / −0.00071
0.10.25−0.0001 / −0.00010−0.0001 / −0.00003−0.0020 / −0.00007
0.10.5−0.0034 / −0.00021−0.0052 / −0.00019−0.0043 / −0.00003
0.20.125−0.0019 / −0.00001−0.0032 / −0.00016−0.0001 / −0.00078
0.2 (default)0.250.0000 / 0.000000.0000 / 0.000000.0000 / 0.00000
0.20.5−0.0022 / −0.00008−0.0076 / −0.00019−0.0057 / −0.00005
0.40.125−0.0056 / +0.00046+0.0016 / +0.00007−0.0020 / −0.00124
0.40.25−0.0028 / +0.00025+0.0014 / −0.00007−0.0018 / +0.00002
0.40.5−0.0057 / −0.00006−0.0100 / −0.00046−0.0109 / −0.00006

Across the scenes and exposures, the maximum absolute deviation from the default is 0.0109 dB in PSNR and 0.00124 in SSIM. This indicates minor sensitivity of the online optimization hyperparameter across different lighting conditions.


Q4-2: Uncertainty-adaptive scheduling of online prior refinement. Motivated by this insightful suggestion that adaptive refinement with uncertainty local guidance, we additionally test spatially adaptive weights derived from uncertainty map between consecutive decoder predictions on 15 LOL images with 28 sampling steps. We retain the initial-prior anchor, τ = 0.25, and a spatial mean weight of α = 0.2, while fixing the checkpoint and initial noise.

Refinement weightsPSNR ↑SSIM ↑
Uniform (default)19.7850.7956
Temporally adaptive spatial weights19.6420.7890

This adaptive variant does not outperform uniform refinement on the tested subset, supporting retention of the default for this setting. The result concerns one temporal-disagreement proxy; it does not rule out other uncertainty-based policies. Selecting the global α or τ per image remains untested.


Q5 / Weakness 2: Near-zero observations and warnings for unreliable content.

Thank you for highlighting this issue. We agree and observe in Fig.11 that faithful reconstruction cannot be guaranteed when observations contain almost no scene information. Our framework can provide a confidence warning based on the calculated uncertainty map to improve the reliabilty of downstream safety-critical applications. The visualization is provided in Appendix Fig.23. For future work, we are going to involve more raw-domain priors, external modality (events or infrared images) and multi-frame signals for unreliable generated content.

AuthorsOpenReview ↗

Official Response to Reviewer 83b1 (Part 3/3)

References

  1. Chen Wei et al. Deep Retinex Decomposition for Low-Light Enhancement. BMVC, 2018. LOL dataset.
  2. Wenjing Wang et al. Zero-Reference Low-Light Enhancement via Physical Quadruple Priors. CVPR, 2024.
  3. Haofeng Huang et al. QuadPrior++: Multi-Dimension Augmented Physical Prior for Zero-Reference Illumination Enhancement. IEEE TPAMI, 48(3):3383–3398, 2026; first published online in 2025.
  4. Chen Chen et al. Learning to See in the Dark. CVPR, 2018. SID dataset.
  5. Jingfeng Yao et al. Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models. CVPR, 2025.
Reviewer_4oydOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents UG-Prior, a framework for generalizable zero-shot low-light image enhancement. It constructs robust and informative priors through uncertainty-guided multi-scale fusion and reliability-aware chromatic encoding. An auxiliary decoder regularizes the generative features by reconstructing clean intrinsic representations, while online prior image co-optimization refines the conditioning prior using reconstruction feedback during inference. However, I have concerns about the separation from the QuadPrior family, and the defination of zero shot.

Critical strengths
  1. The paper presents a coherent approach to improving both the construction and utilization of physical priors under low-light degradation. 2. The combination of uncertainty-guided fusion, chromatic reliability weighting, and prior refinement is technically motivated.
Critical weaknesses

The proposed method shares substantial conceptual similarities with the QuadPrior family, which limits its methodological novelty.

The definition of zero-shot in this paper is questionable. The proposed method requires training on a subset of SA-1B rather than performing direct inference on the test images without additional training. This raises concerns about whether the comparisons reported in Table 1 are fair and valid. Furthermore, the proposed method is not compared with QuadPrior or QuadPrior++ under the same training data conditions, making it difficult to assess its actual performance advantages over these closely related methods.

Questions for authors

1.Although Appendix C explains why practical degradations can violate ideal illumination invariance, the aggregate results in Fig. 3 do not show how prior estimation errors change with increasing darkness and noise. Nor do they directly verify whether the confidence map can identify regions where prior extraction is unreliable. I suggest that the authors evaluate the same set of images at several brightness and noise levels, plot the prior errors against these two factors, and group pixels or patches into confidence bins to report the average prior error in each bin. These results would provide more intuitive support for the authors’ argument.

2.I have concerns about the paper’s zero-shot claim. The title and overall presentation describe UG-Prior as a zero-shot method, but the student model introduced to reduce inference costs requires 8,000 real low-light training images from LOL and LSRW. This does not satisfy the paper’s own definition of zero-shot as requiring no real low-light training data. The authors do not explicitly state that this claim applies only to the teacher. They should clearly distinguish the training settings of the teacher and student and qualify the overall zero-shot and practical application claims accordingly.

3.This paper shares a similar image restoration approach with QuadPrior++ and explicitly adopts its SFT-based prior injection and student architecture. Although the paper compares against QuadPrior and includes an experiment replacing the conditioning prior with QuadPrior, it lacks comparisons with the teacher and student models of the most directly related method, QuadPrior++. Given this close methodological connection and the availability of public code and student weights, I think adding direct comparisons with the QuadPrior++ teacher and students would make the paper’s novelty more convincing.

5 comments
AuthorsOpenReview ↗

Official Comment by Authors (Part 1/4)

We thank the reviewer for highlighting the need to clarify our contributions beyond QuadPrior, distinguish the teacher’s zero-shot setting from student distillation, and provide comparisons under matched training conditions.

Weakness 1: Contribution beyond QuadPrior.

We appreciate the opportunity to clarify the distinction from QuadPrior [1]. Both methods use illumination priors for zero-shot low-light enhancement. QuadPrior offers a valuable physical prior, but its effectiveness depends on obtaining a reliable, informative condition from degraded inputs. This exposes two gaps: prior estimation under real low light and the mismatch between a fixed input-derived condition and the clean intrinsic representation needed for faithful generation.

To address the first gap, the assumption of ideal illumination invariance [1] does not guarantee reliable estimation under real low-light degradation. We improve the robustness–informativeness trade-off through uncertainty-guided multi-scale fusion, adapting the contribution of different scales to local observation reliability. Figure 3 demonstrates the resulting balance between cross-illumination stability and scene-information preservation. The confidence–error analysis (Q1) and frozen-prior reconstruction study (Q3) further support the reliability modeling and information retention, respectively. Furthermore, under the same FLUX backbone and training settings, replacing our prior with QuadPrior’s reduces PSNR from 21.29 to 19.10 dB (Table 3), supporting the contribution of our prior design.

To address the second gap, our contribution lies in the prior-conditioning framework built around FLUX, including clean-prior decoding regularization and test-time prior–image co-optimization. The auxiliary decoder regularizes intermediate generative features toward clean intrinsic representations, while its predictions refine the conditioning prior during sampling rather than leaving the initial estimate fixed. The component ablations in Sec. 4.5 evaluate these mechanisms, and the matched-training-data comparison (W3) further supports the effectiveness of the overall framework.


Weakness 2-1: Zero-shot clarification and comparison.

Our main teacher uses no real low-light training images, following the normal-light-only setting of QuadPrior [1]. This usage is explicit in QuadPrior++ [2, 7], which describes enhancement trained only on normal-light data as zero-shot. Zero-shot enhancement also includes methods without test-time optimization, such as MR. Illuminate [3]. We have clarified this in the main text (Sec.4.4). Thank you for this constructive suggestion!

Table 1 makes the data requirements explicit by listing each method's training data or test-image adaptation setting. QuadPrior++ primarily focuses on improving distilled models that require real low-light training data and therefore fall outside our zero-shot setting. We thus exclude these distilled models from the main quantitative comparison in Table 1 and include them in the distillation comparisons in Q3. On LOL, our UG-Prior achieves 21.29 dB, exceeding QuadPrior's 20.31 dB and MR. Illuminate's 20.94 dB by 0.98 dB and 0.35 dB, respectively. Q2–Q3 further examine our practical value.


Weakness 2-2: Comparison with QuadPrior under matched training data.

We retrain QuadPrior [1] to convergence on the same SA-1B subset [4] used to train UG-Prior for fair comparison. QuadPrior++ makes no modification on the teacher model, so we present matched comparison in Q3 in terms of distillation paradigm.

On LOL/LSRW, UG-Prior increases PSNR by 0.61/1.26 dB, SSIM by 0.046/0.007, and NIMA by 0.022/0.035, while reducing LPIPS by 0.072/0.111. These results support the superiority of our method under matched training data. NIMA [6] evaluates perceptual quality. Bold denotes the best value for each metric within each dataset.

DatasetMethodTraining dataPSNR ↑SSIM ↑LPIPS ↓NIMA ↑
LOLQuadPriorCOCO20.310.8080.2024.043
LOLQuadPriorSA-1B subset20.680.7920.2314.205
LOLUG-PriorSA-1B subset21.290.8380.1594.227
LSRWQuadPriorCOCO16.960.5640.4084.615
LSRWQuadPriorSA-1B subset17.250.5680.4124.886
LSRWUG-PriorSA-1B subset18.510.5750.3014.921
AuthorsOpenReview ↗

Official Comment by Authors (Part 2/4)

Question 1: Does confidence identify unreliable prior estimates?

We provide additional quantitative analysis to verify whether confidence identifies unreliable prior estimates on 291 short–long pairs from the same 49 SID Fuji scenes [5], captured at 33/40/100 ms. We define empirical prior error EE as the patch-normalized MSE between short- and long-exposure priors. Lower error indicates greater reliability. For each Gaussian scale σ\sigma and patch BB, we pair this error with the mean confidence over the same patch. More details are reported in Appendix A.1.

Here, we consider only the scale-dependent prior Pσ=[H,log⁡S,Pstr]P_\sigma=[H,\log S,P^{\rm str}], and ZBZ_B standardizes each patch and channel independently, with ϵ=10−5\epsilon=10^{-5} added to the variance. Confidence qq is computed from local luminance and the structure tensor using Eq. 18. We evaluate 8 × 8, 16 × 16, and 32 × 32 patches at five Gaussian scales. We compute within-image correlations and summarize mean errors in ten confidence-ranked bins. A plot figure represent direct comparison is added in Appendix A.1.

As shown in the following table, fine-scale priors show a stronger negative confidence–error correlation than coarse-scale priors. Coarse-scale priors have lower mean errors and weaker correlations, consistent with lower sensitivity to degradation. The table reports absolute value of Pearson rr for σ=0.5\sigma=0.5 / σ=2.5\sigma=2.5. These results support Eq. 3 and our core motivation. Coarse-scale priors maintain lower mean error even in low-confidence bins, extending useful guidance to regions where fine-scale estimates are less reliable. High-confidence regions can retain fine-scale cues to preserve detail. This allocation improves the empirical robustness–informativeness trade-off.

Patch size33 ms40 ms100 ms
8 × 80.319 / 0.1350.257 / 0.0580.459 / 0.089
16 × 160.446 / 0.2010.349 / 0.0800.582 / 0.116
32 × 320.499 / 0.2590.390 / 0.0950.612 / 0.138

Question 2: The distilled student and the zero-shot claim.

We agree that the distilled student should not be described as zero-shot. Following QuadPrior's data-distillation strategy [1], the original student uses 8,000 real low-light images with teacher-generated targets. We have revised the claim to apply only to the main teacher and make the student's training setting explicit. Q3 presents the supporting quality and efficiency results, together with additional distillation settings.

AuthorsOpenReview ↗

Official Comment by Authors (Part 3/4)

Question 3: Distillation settings and comparison with QuadPrior++.

To complement Q2, we compare against distilled QuadPrior [1] and QuadPrior++ [2] across three distillation settings. Within each setting, the compared methods use the same training data and protocols. The data sources are a normal-light subset of SA-1B [4] (8,000 images) and the real low-light LOL+LSRW training set (8,000 images). UG-Prior adopts the same student architecture and prior injection module as QuadPrior++ [2] for a fair comparison.

[Setting 1: Used in the main text] Teacher-generated distillation. We train students with and without prior injection using teacher-generated enhanced images as targets. Prior injection increases LOL/LSRW PSNR by 0.66/0.30 dB and SSIM by 0.015/0.028, with 0.027 GB of additional peak memory. Relative to the teacher, the student with prior injection reduces FLOPs by a factor of approximately 3.28 × 10⁴ and peak memory by 98.7%, demonstrating a substantial reduction in inference cost while retaining useful enhancement performance. With and without spatial feature transform (SFT) prior injection [2], UG-Prior students outperform QuadPrior students, demonstrating better practical performance and stronger teacher capabilities.

The LOL and LSRW cells report PSNR ↑ (dB) / SSIM ↑ at the original GT resolution. Bold denotes the best quality score within the teacher and student groups separately:

ModelLOLLSRWFLOPs ↓Peak memory (GB) ↓
QuadPrior teacher20.31 / 0.80816.96 / 0.5641.8 × 10¹⁴30.64
UG-Prior teacher21.29 / 0.83818.51 / 0.5754.6 × 10¹⁵32.03
QuadPrior student w/o prior18.62 / 0.76617.39 / 0.5217.750 × 10¹⁰0.385
QuadPrior student w/ SFT prior19.03 / 0.79218.28 / 0.5611.217 × 10¹¹0.412
UG-Prior student w/o prior18.81 / 0.78818.01 / 0.5457.750 × 10¹⁰0.385
UG-Prior student w/ SFT prior19.47 / 0.80318.31 / 0.5731.403 × 10¹¹0.412

[Setting 2: Zero-shot] Frozen prior-to-RGB reconstruction. We train students to map the frozen UG-Prior or QuadPrior representation directly to RGB, using only the same normal-light source images. This zero-shot probing setting tests how much image information the prior retains for reconstruction. The higher reconstruction scores of UG-Prior support the information-retention aspect of our prior design under the same normal-light-only training setting.

The LOL and LSRW cells report PSNR ↑ (dB) / SSIM ↑ at the original GT resolution. Bold denotes the best value for each quality metric.

ModelLOLLSRWFLOPs ↓Peak memory (GB) ↓
Distilled frozen QuadPrior → RGB18.73 / 0.72016.15 / 0.4798.554 × 10¹⁰0.410
Distilled frozen UG-Prior → RGB19.58 / 0.75916.61 / 0.5291.042 × 10¹¹0.410

Comments on the paper

1 comment

Dear AC, this article has a decent level of innovation, with sufficient content and complete experiments. I have ample research experience in this field and I believe that this article is well above the acceptance line of ICLR, so I strongly recommend accepting this article.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.