acceptodds
Under review as a conference paper at ICLR 2027

Light Attention Lights Your World

Abstract

Low-light image enhancement remains challenging under spatially non-uniform illumination. Existing unified restoration pipelines often over-enhance bright areas or insufficiently restore severely underexposed regions, while processing highresolution images incurs substantial computational overhead. To address these challenges, we propose Net, a **L**ightweight **L**ight-Attention Network for **L**ow-**L**ight Image Enhancement. Net performs restoration in the low-resolution feature space using two structurally identical illumination-aware branches that adaptively restore differently illuminated regions. Light-Attention (LA) blocks, the core components of the lightweight backbone, integrate global illumination modeling and local feature interactions. A light-guided learning strategy further encourages branch specialization. A super-resolution reconstruction module recovers fine details from compact feature representations. Extensive experiments on multiple low-light benchmarks demonstrate that Net achieves state-of-the-art performance with substantially lower computational cost. Deployment on an embedded platform further demonstrates its practicality for resource-constrained low-light vision applications. Code will be released upon publication.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low-light image enhancement, Spatially non-uniform illumination, Dual-branch decoupling, Retinex theory

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_2o1AOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes L4 Net, a lightweight low-light image enhancement network addressing both spatially non-uniform illumination and the efficiency of high-resolution inputs. Enhancement is performed in a low-resolution feature space, two branches specialize on bright and dark regions through an adaptive luminance mask, lightweight LA blocks model global illumination via channel-domain attention with local convolution refinement, an SR module recovers high-frequency details, and CLIP text prompts provide semantic guidance.

Critical strengths

The efficiency design is the most convincing strength. Enhancing in a low-resolution feature space and deferring high-resolution computation to SR reconstruction cuts peak GPU memory by 64.2 percent at 1600 by 1200 output with large reductions in FLOPs and inference time, and the deployment on an Orange Pi 5 confirms its practicality for resource-constrained scenarios.

Critical weaknesses
  1. The mask source inconsistency between training and inference is the primary concern. The mask is computed from the bright branch output during training, but that output does not exist at inference, and the paper does not state how the mask is obtained at inference, which directly affects reproducibility and the credibility of the deployment description.
  2. The fairness of the downsampling comparison is questionable. At 4x, L4 Net's SR module is designed for resolution recovery while baselines are evaluated after nearest-neighbor upsampling, putting them at a disadvantage, so a baseline equipped with a comparable SR module is needed.
  3. The novelty is primarily combinatorial. The dual-branch architecture exists in IDB, the channel attention inherits from Restormer, the gamma correction and SR module come from prior work, and CLIP prompt guidance has precedents, so the contribution is an efficient combination of existing components rather than a new concept.
  4. The incremental gain of the new component is limited. Table 2 shows that adding light-guided learning alone drops PSNR from 24.13 to 23.88, and the final improvement comes mainly from the dual-branch and SR combination, so the attribution needs to be clearer.
Questions for authors
  1. How is the mask computed at inference, and is the source consistent with training or based directly on the input image?
  2. Can you add a baseline equipped with a comparable SR module for a fair 4x downsampling comparison?
Reviewer_4diXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes L4 Net for low-light image enhancement. The method uses two independent branches to restore bright and dark regions, with a luminance-based mask to guide branch specialization. The backbone combines channel attention and local convolution, while a super-resolution (SR) module reconstructs high-resolution images from low-resolution features to reduce computational cost.

Experiments on LOLv1, LOLv2-Syn, and ACDC-Night show competitive PSNR and SSIM results. The paper also provides ablation studies, efficiency comparisons, and deployment results on an embedded device.

Critical strengths
  1. The paper addresses a practical problem. Handling non-uniform illumination while reducing the computational cost of high-resolution image enhancement is useful for real-world applications.

  2. The experimental results are promising. L4 Net achieves competitive performance on several benchmarks, especially with downsampled inputs. The efficiency evaluation and embedded-device deployment also show the potential of this method for practical use.

Critical weaknesses
  1. The main empirical concern is that the reported gains are difficult to attribute cleanly to the proposed illumination-guided branch specialization. The method combines dual-branch enhancement, light-guided routing, and a super-resolution reconstruction module, but the current experiments do not fully disentangle which component is primarily responsible for the improvements.

  2. The central design choice of using a hard luminance-based partition is not sufficiently validated. The paper does not study whether the binary routing near the luminance threshold introduces discontinuities or whether softer or multi-level routing strategies would be more appropriate for gradually varying illumination. As a result, the core mechanism is plausible but not yet fully justified.

  3. The paper provides the general training settings and an algorithm outline, but some implementation details are missing, such as the exact LA block configuration, CLIP settings, and preprocessing procedures. Since the code is not currently available, reproducing the reported results may be difficult.

Questions for authors
  1. Why did the authors choose only two branches with hard luminance partitioning? Have soft routing or three/four-branch designs been tested? I am particularly interested in whether hard routing affects the restoration quality of regions near the luminance threshold.

  2. Why is the mask generated using fixed 4×4 regions? Have the authors tested different values of k? Does the optimal partition size change with the input resolution?

  3. Can the authors provide a fairer comparison where competing methods use the same SR reconstruction strategy? This would help clarify whether the performance gain mainly comes from the light-guided dual-branch design or the SR module.

  4. How is branch specialization maintained in the final HR output? Since the SR module directly fuses features from both branches, can the authors provide evidence that the two branches still contribute differently to bright and dark regions?

Reviewer_kk5COpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes L4Net to address the differing enhancement needs of bright and dark regions in high-resolution low-light images and the computational cost of processing these images directly. Two structurally identical branches with independent parameters restore images at low resolution, using learnable gamma correction for preliminary enhancement and modules combining channel attention with local convolutions for feature extraction. During training, the bright branch output generates an adaptive luminance mask that assigns regional supervision to the two branches, while positive and negative text prompts provide auxiliary semantic constraints. The features are then fused, and residual channel attention and upsampling modules reconstruct the high-resolution output. The paper compares restoration quality on three low-light datasets and reports component ablations, GPU memory and runtime measurements, and embedded-device evaluations, demonstrating a trade-off between quality and computational cost. The results support the system's value, but the fairness of comparisons across scales, the evaluation targets for real nighttime scenes, and the independent contribution of luminance-based supervision still require clarification.

Critical strengths
  1. In Table 1, on LOLv1 with 4× downsampling, L4Net improves PSNR from 22.30 to 22.69 dB and SSIM from 0.741 to 0.786 over IDB retrained at each scale, while using fewer than 7% of its parameters.
  2. In Table 4, at the same output size of 1600×1200, the 4× configuration reduces GPU memory from 5.95 to 2.13 GB. The paper also reports embedded-device runtime costs at different resolutions.
Critical weaknesses
  1. Although Section 2.1 discusses IDB, the proposed isomorphic dual branches, dynamic mask-based specialization, and low-resolution restoration followed by super-resolution overlap with it; the main change in regional specialization is replacing difficulty with output luminance. Section 3.3 does not explain why the relative luminance of the enhanced output distinguishes dark-colored content that should be preserved from underexposed regions requiring restoration. Equations (6)–(7) also simply express gradients through fusion with a fixed mask. It therefore remains unclear how changing the basis for specialization constitutes a new methodological contribution.
  2. In Section 4.2, the baselines other than IDB are not retrained at each scale and use nearest-neighbor interpolation to restore the output size, whereas L4Net uses learned super-resolution; Table 4 also compares different input sizes. These differences limit the comparability of the results supporting the claimed quality and efficiency advantages. The same-scale IDB results provide positive evidence, but do not replace comparisons with the other representative methods under matched conditions.
Questions for authors
  1. Why does Section 3.3 construct the luminance mask from the bright branch's enhanced output rather than the input? How does this relative luminance criterion distinguish naturally dark content from underexposed regions, and how does its specialization principle differ substantively from IDB's difficulty-based supervision?
  2. In Section 4.2, is L4Net trained separately at each scale? Would the quality advantages in Table 1 and latency advantages in Table 4 persist if representative baselines used the same training scales and comparable reconstruction?
  3. How are the ACDC-Night reference images obtained, aligned, or synthesized? Please specify the evaluation regions and clarify whether the metrics in Table 1 evaluate restoration using real day/night pairs or restoration of synthetically darkened images.
  4. What new constraints do the fixed exposure prompts and Equation (8) in Section 3.3 introduce beyond CLIP-LIT's existing positive/negative prompt supervision? Do these prompts mainly constrain exposure characteristics, or do they also constrain fidelity to semantic content?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.