acceptodds
Under review as a conference paper at ICLR 2027

Learning Structural Illumination for Unsupervised Low-Light Enhancement

Abstract

Existing unsupervised low-light image enhancement (LLIE) methods often estimate illumination directly from the entire low-light input, without separating its spatially varying illumination pattern, termed relative illumination structure, from the absolute exposure level or preventing unreliable low signal-to-noise ratio regions from biasing the estimate. Moreover, fixed exposure targets impose a scene-agnostic enhancement criterion, limiting adaptation across diverse lighting conditions. Inspired by the spatial propagation of light, we propose a Relative Illumination Structure Estimation (RISE) framework that decouples relative illumination structure from absolute exposure and infers it from reliable bright regions, enabling interpretable and robust enhancement. For scene-adaptive exposure adjustment, we further propose a Dual-Metering Exposure Reference derived from each input, allowing RISE to adapt the enhancement strength to individual scenes and generalize across diverse lighting conditions. Extensive benchmark and real-world generalization experiments show that RISE achieves state-of-the-art performance on most datasets among unsupervised LLIE methods while producing visually natural results. Code will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low-Light Enhancement, Structural Illumination

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_3vNhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents RISE for unsupervised low-light image enhancement. It models spatial illumination adjustments and global exposure separately in log-gain space. Several bright, spatially dispersed patches are selected as Key Illumination Cues, and their brightness differences and distances to other regions are used to build the gain field. For training, the authors introduce DMER, which combines two metering rules to produce an exposure reference from each input image, without paired normal-light supervision.

Critical strengths
  1. RISE is trained only on LSRW-Nikon, yet exceeds the compared prior unsupervised methods in PSNR on six of seven paired benchmarks. It also achieves the highest reported IQA and IAA scores on the five unpaired datasets.
  2. DMER provides training guidance without paired normal-light images or external learned priors.
Critical weaknesses
  1. In the figure, features from the original image are used to predict the KIC target luminance, which is then fed into the structural branch; however, Appendix A.6 states that the relative branch receives only normalized information.

2 . The paper uses a default patch size of 75 × 50 pixels, but does not clarify whether this size applies to the original images or to images resized to a common resolution, nor whether it is adjusted according to the input resolution.

  1. It is suggested that the training dataset, evaluation dataset, and data splits be specified in the ablation section or table captions.

  2. DMER directly generates an enhanced reference image and uses it to constrain the output luminance, patch gains, and KIC luminance of RISE, but Table 9 shows that RISE† achieves higher PSNR than this reference. This result is not contradictory, but the current experiments do not sufficiently explain where the additional gains come from.

  3. Eq. (1) adds epsilon only to the denominator. Therefore, when the mean luminance of a patch is P(p) = 0, log 0 still occurs, and the output is not guaranteed to be finite.

Questions for authors
  1. Under what conditions does the bright-region proxy fail? Please group the cases by saturation ratio, specular reflections, noise in dark regions, and dynamic range, and compare reliability-based selection strategies.

  2. Do the image features at the KICs and the predicted target luminance in Fig. 3 introduce absolute exposure information into the structural branch? How are they normalized?

  3. Which statistics are included in c_scene, and what interpolation method does Up use?

  4. What are the architectures of the image encoder and the illumination encoder, respectively?

  5. The definition of Γ(·) and the value of τ_abs in Eq. (6) are not provided.

8 comments
AuthorsOpenReview ↗

[1/2] Response to Critical Weaknesses

We appreciate your valuable feedback. We hope the following clarifications and additional results address your concerns. W = Critical Weaknesses, Q = Questions For Authors.

To W1:

Thank you for pointing out this ambiguity. We intended to state that relative illumination relationships are modeled using normalized information from the relative illumination map, without directly receiving original image features or global luminance statistics. The KIC target luminance predicted from original image features is still provided to the structural branch as anchor information, so this branch does not rely solely on normalized information. We have revised lines 811 to 813 of the revised manuscript and marked the changes in red.

To W2:

Thank you for pointing out this lack of clarity. All input images are processed at their original resolutions without resizing to a common resolution. The patch size is fixed at 75 × 50 pixels across all datasets and input resolutions and is not adjusted according to image resolution.

To W3:

Thank you for the suggestion. In the original submission, we listed the training datasets used by all compared methods in Appendix C.1, Table 10, due to limited table space. Following your suggestion, we have added details of the evaluation datasets and data splits in lines 972 to 974 of this section in the revised manuscript to further improve the clarity and reproducibility of our experiments.

To W4:

DMER provides soft exposure supervision rather than an upper bound on enhancement quality. RISE does not simply replicate the handcrafted reference image. Instead, joint optimization with DMER exposure supervision, relational modeling based on KIC anchors, and multiple regularization objectives enables the network to learn spatially varying illumination corrections and achieve better enhancement than the reference itself.

Existing ablation experiments further clarify the sources of these gains. On LSRW-Nikon, Table 9 reports a PSNR of 16.297 dB for the DMER reference, while full RISE achieves 16.71 dB. Table 4(c) shows that removing the relative illumination structure or scale equivariance regularization reduces PSNR to 15.16 dB and 16.04 dB, respectively, both below the DMER reference. Table 4(a) also supports the effectiveness of each relational modeling term. Together, these results support the explanation that relative illumination modeling and regularization jointly contribute to the gains beyond the handcrafted reference.

To W5:

Thank you for pointing out this issue. Our implementation already includes the numerical stabilizer ϵ\epsilon in both the numerator and denominator. However, due to an oversight related to experience during manuscript preparation, we omitted the stabilizer from the numerator in Eq. (1). We have corrected this error in the revised manuscript.

AuthorsOpenReview ↗

[2/2] Response to Questions for Authors

To Q1:

Following your suggestion, we added experiments grouped by saturation ratio, specular reflections, noise in dark regions, and luminance range. We compared random selection, bright region selection, reliability filtering, and soft weighting. Bright KIC selection achieves the lowest MAE in all groups except the highest noise quartile, where its MAE is only 0.0027 higher than soft weighting. This indicates that reliability weighting provides a modest improvement, while bright region selection remains competitive under the tested high noise conditions.

Table 6. MAE comparison of KIC selection. Lower is better.

GroupCountRandomBrightHardSoft
Sat. ≤1%1290.08310.08300.08500.0842
Sat. >1%60.07070.05960.07190.0705
No specular1100.08740.08650.09140.0906
Specular250.06110.05250.05350.0532
Noise Q1340.09050.07330.07520.0771
Noise Q4340.06820.06990.07080.0672
Range Q1340.08970.08660.08710.0874
Range Q4340.09190.07290.07710.0732

Count denotes the number of images. Sat. denotes the saturation ratio. Noise refers to noise in dark regions, and Range denotes the luminance range. Q1 and Q4 denote the lowest and highest quartiles. Hard and Soft denote filtering and weighting using reliability scores, respectively.

To Q2:

KIC target luminance is jointly predicted from original image features and relative illumination features. The relative illumination representation is obtained through mean normalization as defined in Eq. (1). The structural branch receives the normalized relative illumination representation and the predicted KIC target luminance, without directly receiving original image features or raw luminance values. KIC target luminance represents the desired luminance after enhancement and provides a target reference for predicting the coefficients of propagation relationships. Therefore, using this target is not equivalent to directly introducing the original exposure level into structural modeling. Scale equivariance regularization further promotes the separation of structure and exposure.

To Q3:

The scene statistics vector cscenec_{\mathrm{scene}} contains three statistics, namely the logarithm of the mean relative luminance of the three KIC anchors, the standard deviation of log relative luminance across the entire image, and the range of log relative luminance among the three KIC anchors. Up\mathrm{Up} uses bilinear interpolation.

To Q4:

The image encoder and illumination encoder share the same architecture but have independent parameters. The image encoder takes a three channel RGB image, while the illumination encoder takes a single channel relative log luminance map ZZ. Both encoders contain three stages with 16, 32, and 64 channels, respectively. Each stage consists of a double convolution block and a residual block, using 3×33\times3 convolutions and LeakyReLU. The last two stages downsample through convolutions with a stride of 2. The first stage produces local features, while the final stage output undergoes global average pooling to obtain global features.

To Q5:

Γ(⋅)\Gamma(\cdot) takes the minimum of the target exposure gain and the highlight protection gain 0.92/max⁡(Y99,ϵ)0.92/\max(Y_{99},\epsilon), where Y99Y_{99} denotes the 99th percentile of the input luminance. This keeps the 99th percentile luminance of the absolute exposure reference at or below 0.92, leaving headroom below the saturation value of 1. We set the absolute exposure target to τabs=0.36=2m\tau_{\mathrm{abs}}=0.36=2m, where m=0.18m=0.18 is the middle gray reference. This choice is consistent with the luminance key values used by Reinhard et al. and provides a luminance target above the middle gray reference for low light enhancement.

Dear Reviewer 3vNh,

I hope this message finds you well. As the discussion period is nearing its end with less than four days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we’re eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Reviewer_mAEgOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RISE, an unsupervised low-light image enhancement framework that estimates relative illumination structure from reliable bright regions while decoupling it from absolute exposure control. It further introduces a Dual-Metering Exposure Reference (DMER) to enable scene-adaptive enhancement by integrating global exposure information with spatial illumination cues.

Critical strengths

Extensive experiments across multiple benchmarks demonstrate strong performance among unsupervised methods, with favorable computational efficiency and promising cross-dataset and real-world generalization capabilities.

Critical weaknesses

(1) DMER essentially reduces to a global exposure gain followed by Gaussian smoothing, raising questions about its ability to model spatial illumination and whether the performance gains primarily stem from the hand-crafted exposure reference.

(2) The scene-adaptive exposure targets of DMER may conflict with the scale-equivariance regularization, raising concerns about the consistency of the training objectives and the claimed exposure invariance.

(3) The KIC-based relational field can be reformulated as a spatially varying gamma-like correction, making its novelty and advantages over simpler intensity-mapping methods unclear.

Questions for authors
  1. Could the authors further clarify how DMER captures spatial illumination structure beyond global exposure adjustment and whether RISE offers additional benefits over the hand-crafted reference?

  2. Could the authors elaborate on how DMER’s scene-adaptive exposure targets align with the scale-equivariance objective?

  3. Could the authors further clarify the advantages of KIC-based relational modeling over simpler gamma-like intensity mappings, potentially supported by additional comparisons?

12 comments

We appreciate your valuable feedback. We hope the following clarifications and additional results address your concerns. W = Critical Weaknesses, Q = Questions For Authors.

To Q1 & W1:

DMER uses spatial illumination information to determine an exposure reference for each image. The relative meter first performs confidence weighted local aggregation and normalization at each location, then computes global statistics of the resulting light field. The exposure decision therefore depends on the spatial arrangement of bright and dark regions. Appendix Table 8 directly verifies this property. Permuting image patches while preserving the luminance distribution and mean produces average changes ranging from 11.14% to 41.77% in the metering statistic across datasets. This demonstrates that DMER is sensitive to spatial illumination layout, while RISE further learns spatially varying regional gains.

Regarding the contribution of the handcrafted reference, Appendix Table 9 directly evaluates the DMER reference as the enhanced output. DMER itself provides a strong baseline, but RISE† further improves PSNR on all five datasets by 0.413 to 1.318 dB. These results show that, with DMER exposure supervision and multiple regularization terms, RISE learns relative illumination structure and exposure control that yield better enhancement than the reference alone.

To Q2 & W2:

We clarify that DMER and scale equivariance regularization play complementary roles. DMER generates an exposure target T(I)T(I) only for the original input II to supervise the enhanced output. The scaled input sIsI participates only in auxiliary scale equivariance regularization, without generating or imposing an additional DMER exposure target.

Specifically, DMER provides an adaptive exposure reference for each scene, while scale equivariance regularization constrains the model response to input exposure changes within the same scene. This regularizer encourages stable relative illumination parameters through Θs≈Θ\Theta_s \approx \Theta and compensates for input brightness changes through ηs≈η−log⁡s\eta_s \approx \eta - \log s, promoting approximately consistent enhanced outputs. This constraint allows different scenes to have different exposure targets and does not require the model to simultaneously satisfy two potentially different targets T(I)T(I) and T(sI)T(sI).

Our experiments also support this design. Figure 4 and Appendix Table 7 show the stability of structural parameters and the adaptive changes in exposure offsets. Table 4(c) further shows that including this regularizer improves PSNR from 16.04 dB to 16.71 dB.

To Q3 & W3:

RISE uses reliably illuminated regions as KIC anchors and constructs the gain field from relative luminance and spatial relationships between each region and multiple KICs, together with scene context. Compared with simple gamma mappings, this formulation explicitly incorporates illumination relationships across regions and anchor constraints, providing a structural prior for regional gain estimation.

We additionally compare RISE with different global gamma settings and a spatially varying gamma baseline. As shown in the tables, RISE achieves higher PSNR than all tested global gamma settings and outperforms the spatial gamma baseline in PSNR, SSIM, and LPIPS, supporting the advantages of relational modeling with KIC anchors.

Table 4. Comparison with global gamma correction. PSNR is reported in dB.

γ\gamma0.10.30.50.70.9
PSNR ↑9.2015.7414.7511.729.91

Table 5. Comparison with spatially varying gamma correction.

MethodPSNR ↑SSIM ↑LPIPS ↓
Spatial gamma15.940.4880.288
RISE16.710.4910.236

Dear Reviewer mAEg,

I hope this message finds you well. As the discussion period is nearing its end with less than four days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we’re eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

Dear Reviewer mAEg,

We hope this revision addresses your concerns. Please let us know if there is anything further we can clarify; we would be happy to discuss any remaining issues.

Reviewer_be8rOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RISE, an unsupervised low-light image enhancement framework motivated by the reliable relative illumination relationship inferred from high intensity areas. Specifically, RISE estimates spatially varying illumination patterns using sparse bright regions as high-SNR anchors, and introduces a Dual-Metering Exposure Reference to generate input-adaptive training targets instead of using a fixed scene-agnostic exposure target. Extensive experiments on paired benchmarks and unpaired datasets demonstrate that RISE achieves promising performance among unsupervised LLIE methods with an extremely lightweight architecture.

Critical strengths

1.This work is well motivated and models unsupervised LLIE as a joint estimation of two disentangled components: a relative illumination structure field capturing spatial lighting patterns, and a global scalar offset controlling global exposure level. The introduction of DMER also improves the generalization capability using scene-adaptive enhancement references.

2.The experimental results across several paired and unpaired datasets show a clear margin compared to unsupervised LLIE baselines.

Critical weaknesses

1.The current ablation did not stand out the unique contribution from KIC alone. What is the performance of a strong baseline with DMER module and without KIC anchor design?

2.Related to previous point and mentioned in the limitation, the feasibility of KIC holds when the selection of high-PSNR indeed captures informative and reliable estimation of underlying illumination.Though there are cases of RISE recovering from dark images, some experimental analysis on high-noise low-light datasets are encouraged to be included to see the capability boundary of RISE.

3.More sensitivity analysis of key hyper-parameters are expected as there are many coefficients involved. For example, the number of KIC are selected within [1,5,3], are they proper candidates for images with various sizes?

4.The writing and notation could be further improved. The relation between pipeline in Fig3 and DMER in Fig5 could be better illustrated.

5.Typo: Tab4 . Brightest+DMER isntead of DME‘I’ (R is partially covered)

Questions for authors

Please see the Weaknesses above for details.

1.The current ablation did not stand out the unique contribution from KIC alone. What is the performance of a strong baseline with DMER module and without KIC anchor design?

2.Related to previous point and mentioned in the limitation, the feasibility of KIC holds when the selection of high-PSNR indeed captures informative and reliable estimation of underlying illumination.Though there are cases of RISE recovering from dark images, some experimental analysis on high-noise low-light datasets are encouraged to be included to see the capability boundary of RISE.

3.More sensitivity analysis of key hyper-parameters are expected as there are many coefficients involved. For example, the number of KIC are selected within [1,5,3], are they proper candidates for images with various sizes?

2 comments

We appreciate your valuable feedback and support of our work. We hope the following clarifications and additional results address your concerns. W = Critical Weaknesses, Q = Questions For Authors.

To Q1 & W1:

The core gain formulation of RISE is built around KIC, estimating regional gains through relative illumination relationships to these anchors. Removing KIC thus requires reformulating gain prediction. We evaluate a baseline that directly predicts regional gains under DMER supervision without KIC. Its performance drop in the table supports the key role of KIC in constraining gain estimation through relative illumination relationships anchored in reliable bright regions.

Table 1. Ablation of KIC anchors under DMER supervision.

MethodPSNR ↑SSIM ↑LPIPS ↓
RISE16.710.4910.236
w/o KIC16.180.4860.249

To Q2 & W2:

We additionally evaluate RISE on SID [1], a low light dataset with high noise. As shown in the table, RISE achieves the best PSNR and SSIM among the compared methods, demonstrating competitive performance. However, restoration quality remains below that on datasets with less noise, such as LOL. When noise dominates images under extreme darkness, illumination estimation becomes less reliable across methods, as discussed in our limitations.

[1] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to See in the Dark. CVPR, 2018.

Table 2. Performance comparison on SID.

SupervisionMethodPSNR ↑SSIM ↑
SupervisedUFormer18.540.577
SupervisedRetinexNet16.480.578
UnsupervisedEnGAN17.230.543
UnsupervisedRUAS18.440.581
UnsupervisedCLODE13.700.225
UnsupervisedRISE19.060.583

To Q3 & W3:

We retain the original resolution of all images without resizing, including 104 different resolutions in SICE. We use 3 KICs and a region size of 75 × 50 across all datasets. Tables 1 and 2 in the paper support the applicability of this setting across resolutions. For further validation, we conduct sensitivity analyses of KIC count and region size on four datasets spanning resolutions from 384 × 384 to 3840 × 2160. As shown in the table, performance shows modest variation across the tested settings, and higher resolutions do not necessarily require more KICs, further supporting the robustness of the uniform settings.

Table 3. Sensitivity to KIC count N and region size S. Results are PSNR in dB.

DatasetResolutionN=1N=3N=5S=50×25S=75×50S=150×100
LOL-v1600×40020.0720.2120.4319.9520.2120.08
LOLv2-Synthetic384×38418.5618.7218.5118.7018.7218.61
LSRW-Nikon960×64016.6516.7116.8116.6616.7116.59
UHD-LL3840×216022.3722.2222.0321.9822.2222.30

To Q4 & Q5:

Thank you for pointing these out. We have revised the relevant text, figures, and formatting in the revised manuscript.

Dear Reviewer be8r,

I hope this message finds you well. As the discussion period is nearing its end with less than two days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we’re eager to address any remaining issues to improve our work. Thank you for your time and effort in reviewing our paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.