MirrorDistill: Illumination-Aware Latent Distillation for Efficient Low-Light Restoration
Authors: Farida Mohsen, Tala Zaim, Nurul Izni Rusli, Ali Al-Zawqari, Ali Safa, Samir Brahim Belhaouari
Organizations: College of Science and Engineering, Hamad Bin Khalifa University, Doha, Qatar. · Faculty of Electrical Engineering & Technology, Universiti Malaysia Perlis, Malaysia · ELEC Department, Vrije Universiteit Brussel, Brussels, Belgium.
Low-light image enhancement (LLIE) is an im- portant component of visual sensing systems operating under degraded illumination, including nighttime surveillance, au- tonomous navigation, remote sensing, and inspection in poorly lit industrial environments. Most LLIE methods rely on output- level reconstruction losses that supervise only the final restored image, leaving the intermediate feature recovery process weakly constrained. This paper proposes MirrorDistill, an illumination- aware latent distillation framework that links the low-light and clean domains through feature mirroring. During training, a shared encoder and an exponential-moving-average teacher decoder process the clean reference image to generate clean- domain latent targets. These targets supervise the low-light student at two levels: raw encoder features and standardized multi-scale decoder projections. The alignment is applied layer by layer, while a proposed illumination-aware weighting scheme gives greater emphasis to underexposed regions. The teacher and reference branches are used only during training, so inference requires only the lightweight student encoder-decoder and in- troduces no teacher-side computational cost. Under evaluation on the standard LOL benchmarks, MirrorDistill outperforms the state-of-the-art methods on the real-captured LOL-v2-Real set, while having the lowest compute complexity (GMACs) and while remaining competitive on the LOL-v1 and LOL-v2-Synthetic datasets. Ablation studies further show the contributions of the encoder mirror, decoder mirror, and illumination-aware weighting. Finally, we release our code as open-source for the benefit of future research.
Figures & tables
Fig. 1: PSNR vs. Computational cost (GMACs, log scale) on the LOL-v2-real dataset [ 10 ] .
Fig. 2: Overview of MirrorDistill. A shared encoder SE processes the low-light input (student, blue) and the clean reference (teacher, gray); the teacher decoder TD is an exponential moving average of the student decoder SD . Illumination-aware mirror losses align student to teacher at the encoder ( Lenc , raw features) and decoder ( Ldec , per-sample standardized projections), each weighted by W=1+β(1−Y~) ( β=0.6 ) to emphasize dark regions. The clean branch is stop-gradient and is discarded at inference. The final deployed model during inference is SE+SD only.
GMACs
Params
Inference Time
LOL-v1
LOL-v2-real
LOL-v2-syn
Method
(G)
(M)
(ms)
PSNR
SSIM
PSNR
SSIM
PSNR
SSIM
RetinexNet [ 11 ]
587.47
0.84
–
16.77
0.560
15.47
0.567
17.13
0.798
KinD [ 12 ]
34.99
8.02
–
20.86
0.790
14.74
0.641
13.29
0.578
EnlightenGAN
61.01
114.35
–
17.48
0.650
18.23
0.617
16.57
0.734
Restormer [ 15 ]
144.25
26.13
104 †
22.43
0.823
19.94
0.827
21.41
0.830
MIRNet [ 20 ]
785.0
31.76
205 †
24.14
0.830
20.02
0.820
21.94
0.876
TABLE I: Quantitative comparison and computational complexity. PSNR/SSIM are reported on LOL-v1, LOL-v2-Real, and LOL-v2-Synthetic at full resolution evaluation. Compute complexity is reported as well (when available). Best and second-best quantitative results are shown in bold and underlined, respectively. † Runtime reported in [ 30 ] .
Fig. 3: Qualitative comparison on LOL-v2-real.
Configuration
PSNR
SSIM
Lrec ( 7 ) only
21.81
0.8888
+ encoder mirror
23.15
0.9059
+ decoder mirror
23.31
0.9014
+ both mirrors (MirrorDistill)
24.04
0.9093
MirrorDistill without illumination weighting ( β=0 )
22.81
0.9048
TABLE II: Ablation of the proposed components and their impact on the LOL-v2-real image recovery quality at 256×256 evaluation resolution.
This letter presents a novel training approach and loss function for learning low-light image enhancement auto-encoders. Our approach revolves around the use of a teacher-student auto-encoder setup coupled to a progressive learning approach where multi-scale information from clean image decoder feature maps is distilled into each layer of the student decoder in a mirrored fashion using a newly-proposed loss function termed Illumination-Aware Mirror Loss (IAML). IAML helps aligning the feature maps within the student decoder network with clean feature maps originating from the teacher side while taking into account the effect of lighting variations within the input images. Extensive benchmarking of our proposed approach on three popular low-light image enhancement datasets demonstrate that our model achieves state-of-the-art performance in terms of average SSIM, PSNR and LPIPS reconstruction accuracy metrics. Finally, ablation studies are performed to clearly demonstrate the effect of IAML on the image reconstruction accuracy.
Farida Mohsen, Tala Zaim, Ali Al-Zawqari +2
College of Science and Engineering, Hamad Bin Khalifa University, Doha, Qatar. · ELEC department, Vrije Universiteit Brussel, Brussels, Belgium.
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
Low-light image enhancement (LLIE) remains challenging for lightweight models because illumination restoration and color fidelity are difficult to optimize simultaneously in the RGB color space. Although recent color-decoupled methods separate luminance and chrominance representations, they primarily optimize luminance as an enhancement target, leaving its potential as an explicit guidance prior largely unexplored during feature reconstruction. To address this limitation, we propose IDM-Net, a lightweight Illumination-Decoupled Modulation Network for low-light image enhancement. IDM-Net adopts a dual-encoder architecture consisting of a structure encoder that extracts multi-scale appearance features from the RGB image and a lightweight illumination encoder that learns illumination priors from the decoupled luminance (Y) channel. To effectively exploit these priors, we introduce an Illumination-Guided Modulation (IGM) module that injects multi-scale illumination cues into the decoder through spatially adaptive affine modulation, enabling accurate brightness restoration while preserving natural color consistency. Furthermore, we design a lightweight Feature Refinement Block (FRB) to progressively suppress degradation artifacts and recover fine-grained image details during reconstruction. Extensive experiments on multiple standard low-light image enhancement benchmarks demonstrate that IDM-Net achieves competitive performance among lightweight LLIE methods while maintaining an excellent balance between restoration quality and computational efficiency.
Cheng-Yen Hsiao, Jing-Ming Guo
Department of Electrical Engineering National Taiwan University of Science and Technology Taipei, Taiwan