Towards Controllable Low-Light Image Enhancement: A Continuous Multi-illumination Dataset and Efficient State Space Framework
Authors: Hongru Han, Tingrui Guo, Liming Zhang, Yan Su, Qiwen Xu, Zhuohua Ye
Organizations: Department of Computer and Information Science, University of Macau, Taipa, Macau 999078, China · School of Automation, China University of Geosciences, Wuhan 430074, China · Department of Electromechanical Engineering, University of Macau, Taipa, Macau 999078, China
Low-light image enhancement (LLIE) has traditionally been formulated as a deterministic mapping. However, this paradigm often struggles to account for the ill-posed nature of the task, where unknown ambient conditions and sensor parameters create a multimodal solution space. Consequently, state-of-the-art methods frequently encounter luminance discrepancies between predictions and labels, often necessitating "gt-mean" post-processing to align output luminance for evaluation. To address this fundamental limitation, we propose a transition toward Controllable Low-light Enhancement (CLE), explicitly reformulating the task as a well-posed conditional problem. To this end, we introduce CLE-RWKV, a holistic framework supported by Light100, a new benchmark featuring continuous real-world illumination transitions. To resolve the conflict between luminance control and chromatic fidelity, a noise-decoupled supervision strategy in the HVI color space is employed, effectively separating illumination modulation from texture restoration. Architecturally, to adapt efficient State Space Models (SSMs) for dense prediction, we leverage a Space-to-Depth (S2D) strategy. By folding spatial neighborhoods into channel dimensions, this design allows the model to recover local inductive biases and effectively bridge the "scanning gap" inherent in flattened visual sequences without sacrificing linear complexity. Experiments across seven benchmarks demonstrate that our approach achieves competitive performance and robust controllability, providing a real-world multi-illumination alternative that significantly reduces the reliance on gt-mean post-processing.
Figures & tables
Fig. 1: Comparison between static “one-to-one” benchmarks (top) and our Light100 dataset (bottom). While traditional datasets are limited to a single ground-truth mapping, Light100 captures 100 progressive lighting levels. This allows for a flexible mapping governed by the target luminance level β , accommodating various user preferences and ambient conditions.
Fig. 2: Visual comparison of state-of-the-art deterministic enhancement methods on the LOL-v2-Real dataset. Most advanced models produce outputs with a higher average luminance than the Test Ground Truth (GT), demonstrating a systematic luminance mismatch that stems from the deterministic mapping paradigm.
Fig. 3: Quantitative analysis of evaluation bias on the LOL-v2-Real dataset. (a) Luminance distribution analysis: outputs of SOTA methods tend to align with the global Training GT mean rather than adapting to the specific illumination of individual test samples. (b) The ”gt-mean” heuristic is commonly employed to evaluate the fidelity of texture and color by neutralizing the global luminance gap, leading to substantial PSNR gains. Here, w/o and w/ denote evaluations without and with this post-hoc adjustment, respectively.
Fig. 4: Overview of the Proposed CLE-RWKV Framework. (a) The overall pipeline operates in the decoupled HVI color space, where target luminance commands ( β ) modulate features via FiLM. (b) The S2D-Net Backbone employs a split-transform-merge strategy, where the L-RWKV Block serves as the core processor. (c) Analysis of the Periodic Shuffle Strategy. Left: The PS-DS Embedding layer (Pixel Shuffle Downsampling with r followed by 1×1 Conv) demonstrates that semantic topology and features are preserved even at lower resolutions. Right: Visualization of the implicit 2D Token Shift. Compared to conventional 1D flattening, our periodic shuffle strategy folds spatial neighborhoods into the channel dimension, effectively bridging the “scanning gap” and enabling the L-RWKV block to capture local details while mitigating long-range forgetting.
Fig. 5: Resolving the Intensity-Chroma Conflict. (Left) Real physical captures at low illumination suffer from diminished SNR, manifesting as both stochastic noise and severe color shifts. (Middle) The hybrid target synthesizes the target-level intensity ( Imax ) with the pristine chromaticity and texture ( H^,V^ ) extracted from the high-quality reference (Right). This strategy provides a noise-free, color-accurate guidance that decouples illumination control from intrinsic degradation restoration.
Method
LOL-v1
LOL-v2-Real
LOL-v2-Syn
Light100
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
LLFlow+gt-mean [ 9 ]
25.18
0.873
0.114
25.63
0.876
0.159
26.82
0.949
0.043
22.42
0.799
0.230
PyDiff+gt-mean [ 10 ]
27.19
0.882
0.109
27.10
0.881
0.150
26.63
0.929
0.070
21.83
0.805
0.191
LLFormer+gt-mean [ 11 ]
26.30
0.833
0.163
27.61
0.856
0.173
29.50
0.948
0.045
25.71
0.863
0.121
GSAD+gt-mean [ 12 ]
27.84
0.877
0.091
28.81
0.895
0.095
28.67
0.944
0.047
23.70
0.841
0.142
CIDNet+gt-mean [ 13 ]
28.20
0.889
0.079
28.14
0.892
0.101
29.56
0.950
0.047
25.48
0.869
0.087
TABLE II: Quantitative Comparison of CLE Paradigm. We compare our method with state-of-the-art LLIE approaches on four benchmarks: LOL-v1 [ 21 ] , LOL-v2-Real [ 23 ] , LOL-v2-Syn [ 23 ] , and our Light100. Note that all baseline methods were retrained on the Light100 dataset for fair comparison. Best and second-best results are highlighted.
Fig. 6: Visual comparisons on LOL-v1 (first row), LOL-v2-Real (second row), and LOL-v2-Syn (third row). The region within the red box is zoomed in and shown at the bottom left corner.
Fig. 7: Images of the same scene at different luminance levels are presented in rows two, four, and six, with the corresponding luminance histograms (red) positioned above each image. For a clearer comparison, the GT histogram (blue) is overlaid onto the corresponding histograms in each row. While linear scaling tends to amplify noise and introduce distribution bias into deterministic baseline models, the proposed CLE-RWKV method achieves accurate histogram alignment and superior perceptual fidelity.
Fig. 8: Efficiency Analysis on LOL-v2-Real. Each bubble represents a model, where the solid colored area corresponds to the peak memory footprint (labeled in MB) and the center white dot indicates the precise PSNR-FPS coordinate. All metrics are tested on an NVIDIA RTX 4090 with 3×256×256 input. Our CLE-RWKV (red bubble) achieves the optimal trade-off across all metrics.
Method
Configuration
Speed
LOL-v1
Light100
(fps)
PSNR
SSIM
PSNR
SSIM
Color Space
sRGB
-
27.06
0.871
29.07
0.902
Architecture
UNet
33.4
27.19
0.872
28.59
0.899
S2D ( P=1 )
19.9
28.57
0.879
28.41
0.896
S2D ( P=2 )
60.4
29.03
0.890
28.97
0.903
Loss Func.
ℓ1 Only
-
28.31
0.855
29.18
0.900
TABLE III: Ablation study. We evaluate each component’s contribution. Speed (FPS) is measured on an RTX 4090 with 3×256×256 input. P denotes the patch size. Best and second-best results are highlighted.
Existing deep learning-based low-light enhancement methods are typically trained on limited datasets with single enhancement targets, which restricts their generalization ability and controllability in real-world applications. To overcome these limitations, we propose ControlLight, a controllable, consistent, and generalizable framework for low-light enhancement. We first construct a large-scale dataset of real-world degraded images with continuous illumination-strength supervision. To further ensure consistent outputs under different control strengths, we introduce a misalignment-aware weighted flow matching loss that preserves image structure across continuous enhancement strengths. ControlLight allows users to edit real-world degraded low-light images toward satisfactory enhancement results by flexibly controlling the strength while preserving visual consistency and realism. Extensive experiments show that ControlLight achieves state-of-the-art performance against existing low-light enhancement approaches while demonstrating strong continuous controllability and generalization to real-world scenarios.
Yufeng Yang, Jianzhuang Liu, Jisheng Chu +4
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences · Zhejiang University
Low-Light Image Enhancement (LLIE) has long been a challenging problem in low-level vision, as insufficient illumination often leads to low contrast, detail loss, and noise. Recent studies show that deep learning-based Retinex theory can effectively decouple illumination and reflectance. However, existing methods frequently suffer from over-enhancement or color distortion, and often assume uniform noise or ideal lighting. To address these limitations, we propose InterLight, a novel framework that systematically excavates and operationalizes intrinsic illumination priors for LLIE.Our core insight is that robust enhancement requires not just estimating illumination, but constructing an illumination-aware pipeline. We first inject sensor-level illumination-response priors via physics-guided augmentation, then represent the degradation through adaptive prompts conditioned on the scene's latent illumination state. This explicit representation directly guides a luminance-gated intrinsic memory mechanism to selectively compensate for information loss, prioritizing reconstruction in dark regions while preserving fidelity in bright ones. Finally, the entire process is regularized by a self-supervised consistency objective that distills illumination-invariant features. By deeply exploiting intrinsic illumination priors, our method achieves clearer textures and more visually coherent enhancement results. Extensive experiments across multiple benchmarks demonstrate the effectiveness of our approach. Code is available at: https://github.com/House-yuyu/InterLight.
Ziqi Wang, Xu Zhang, Laibin Chang +3
National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University · Department of Computer Science, University of Macau · Department of Computer Vision, Mohamed bin Zayed University of Artificial Intelligence +1
Images captured by consumer electronic devices, such as mobile phones and digital cameras, often suffer from low-light degradation due to sensor limitations and imaging pipelines, which degrades visual quality and affects downstream vision tasks. Existing methods based on Convolutional Neural Networks (CNNs) and Transformers have dominated current low-light image enhancement (LIE) due to their excellent ability to model hierarchical features. However, CNNs operate in local receptive fields that cannot model long-range dependencies, while Transformers overcome this problem but incur substantial computational costs. To address these challenges, we propose MambaLIE, a Scene Light Intensity-Boosted Low-Light Image Enhancement method based on a State Space Model (SSM). We first introduce scene light intensity to improve the structural distribution of illumination, which is then gated with the low-light input to guide enhancement. To better model the illumination while maintaining computational efficiency, we propose the Locally Enhanced State Space Model (LESSM) for efficient light enhancement. Our LESSM contains two branches: an SSM branch and a Local Enhanced branch, where the former is used to model the long-range dependencies with linear time complexity, while the latter is used to enhance local feature representations. Extensive experiments demonstrate that MambaLIE outperforms state-of-the-art CNN-based and Transformer-based LIE methods on four widely used synthetic benchmarks and five publicly available real-world benchmarks in terms of accuracy, speed, and model size, making it suitable for practical deployment on resource-constrained devices.
Wanshu Fan, Xiangyu Li, Cong Wang +4
National and Local Joint Engineering Laboratory of Computer Aided Design, School of Software Engineering, Dalian University, Dalian, China · Hong Kong Polytechnic University, Hong Kong, China · School of Computer Science and Technology, Dalian University of Technology Dalian, China +1