Organizations: Thrust of Robotics and Autonomous Systems, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou 511453, China · Key Laboratory of Knowledge Engineering with Big Data, Ministry of Education of China, Hefei 230601, China · Innovation School of Artificial Intelligence, Hefei University of Technology, Hefei 230601, China · School of Control Science and Engineering, Shandong University, Jinan 250061, China · School of Computer Science and Technology, Harbin Institute of Technology, Harbin 150001, China · School of Data Science, Lingnan University, Hong Kong
Low-light image enhancement (LLIE) is essential to numerous vision applications operating under adverse illumination. Despite substantial advances, real-world nighttime enhancement remains challenging due to the scarcity of strictly aligned day-night pairs and complex degradations involving extreme darkness and heterogeneous artificial light sources, leading to fragile generalization. To overcome these challenges, we propose MrLIE (MLLM-empowered reference Low-light Image Enhancement), a novel reference-based unsupervised LLIE framework that synergizes the generative and reasoning capabilities of multimodal large language models (MLLMs) with conditional diffusion. MrLIE employs the MLLM as a high-fidelity reference synthesizer and diagnostic critic, iteratively generating, diagnosing, and correcting normal-light references directly from degraded nighttime observations. Based on this, our method governs the enhancement process through two complementary mechanisms, in which a confidence offset module rectifies spatial-geometric inconsistencies for reference reliability and a structure-anchored injection scheme harmonizes aligned reference cues, low-light structural constraints, and ControlNet additive anchoring during the denoising process. Extensive experiments on diverse benchmarks demonstrate that MrLIE achieves superior perceptual quality, natural appearance, and robust generalization.
Figures & tables
Fig. 1: Motivation for MrLIE. Existing LLIE methods remain insufficient for real nighttime restoration, while one-shot daytime references often provide incomplete or spatially shifted guidance. MrLIE addresses these by iteratively constructing daytime references and aligning imperfect reference cues for faithful daytime restoration. The third-part visualization is obtained by subtracting the nighttime input from each corresponding image.
Fig. 2: Overall framework of the proposed MrLIE. It first obtains a high-quality normal-light reference through iterative MLLM-guided generation, then performs aligned-reference diffusion restoration by jointly injecting low-light structure, aligned reference features, and diffusion control signals.
Fig. 3: Diagnosis-driven daytime reference generation. MLLM-Gen produces candidates from the nighttime input, while MLLM-Critic and no-reference image quality assessment jointly diagnose residual degradations and guide prompt refinement until a reliable reference is selected.
Fig. 4: Overview of the proposed Confidence Offset Module (COM) and Structure-Anchored Injection (SAI). COM aligns the generated reference features with the low-light features, while SAI injects reference and structural cues into the diffusion process for faithful enhancement.
Type
Method
Venue
DICM
NPE
NIQE ↓
MUSIQ ↑
BRISQUE ↓
PI ↓
CLIP-IQA ↑
LIQE ↑
NIQE ↓
MUSIQ ↑
BRISQUE ↓
PI ↓
CLIP-IQA ↑
LIQE ↑
S
LLFormer
AAAI’23
3.695
59.314
15.720
3.128
0.502
3.129
3.804
67.587
13.563
2.696
0.516
3.894
RetinexFormer
ICCV’23
4.365
56.604
28.089
3.902
0.503
2.971
4.010
64.493
19.134
3.280
0.458
3.384
DFFN
TGRS’24
3.746
55.429
17.242
3.214
0.397
2.314
3.570
63.620
14.709
2.841
0.316
3.019
HVI
CVPR’25
3.545
57.976
22.226
3.080
0.588
3.143
3.879
67.560
18.153
2.923
0.546
3.698
LBM
ICCV’25
4.813
60.868
26.486
3.953
0.560
3.637
4.485
68.028
18.070
3.214
0.576
4.569
TABLE I: Quantitative comparison on DICM and NPE. Best and second-best results are highlighted in bold and underlined, respectively.
Type
Method
LIME
MEF
NIQE ↓
MUSIQ ↑
BRISQUE ↓
PI ↓
CLIP-IQA ↑
LIQE ↑
NIQE ↓
MUSIQ ↑
BRISQUE ↓
PI ↓
CLIP-IQA ↑
LIQE ↑
S
LLFormer
4.153
58.442
15.460
3.184
0.441
2.690
3.650
59.257
16.815
3.079
0.444
3.037
RetinexFormer
4.667
60.480
25.917
3.753
0.436
2.899
4.272
58.789
29.772
3.616
0.448
3.223
DFFN
3.923
60.924
16.384
2.821
0.410
2.876
3.755
62.567
22.944
2.958
0.445
3.290
HVI
4.163
62.378
20.714
3.172
0.461
3.308
3.697
66.736
24.588
2.854
0.474
3.980
LBM
4.637
65.983
21.902
3.330
0.535
3.798
4.435
64.865
24.377
3.228
0.541
4.068
TABLE II: Quantitative comparison on LIME and MEF. Best and second-best results are highlighted in bold and underlined, respectively.
Type
Method
U3D
iSAID-dark
NIQE ↓
MUSIQ ↑
BRISQUE ↓
PI ↓
CLIP-IQA ↑
LIQE ↑
PSNR ↑
SSIM ↑
LPIPS ↓
NIQE ↓
MUSIQ ↑
BRISQUE ↓
PI ↓
CLIP-IQA ↑
LIQE ↑
S
LLFormer
4.938
38.948
24.583
3.227
0.253
1.341
14.744
0.290
0.988
6.690
39.601
46.091
4.420
0.236
1.016
RetinexFormer
3.788
38.995
20.401
2.850
0.246
1.356
17.521
0.353
0.705
6.931
40.405
34.031
4.417
0.275
1.063
DFFN
5.374
37.678
32.314
3.556
0.264
1.270
14.997
0.265
0.878
9.676
38.719
59.826
6.076
0.276
1.054
HVI
5.370
37.777
37.914
3.338
0.285
1.462
13.301
0.202
1.065
10.715
37.895
64.101
6.523
0.294
1.044
LBM
5.449
58.362
30.503
3.510
0.555
2.871
16.394
0.406
0.476
9.396
55.322
48.769
5.697
0.468
2.446
TABLE III: Quantitative comparison on U3D and the iSAID-dark validation benchmark. Unsupervised methods are trained on U3D where applicable and directly evaluated on iSAID-dark without fine-tuning.
Fig. 5: Qualitative generalization results on DICM and NPE. MrLIE produces natural exposure and color while avoiding excessive enhancement, color distortion, and structural drift under real-world image domain shifts.
Fig. 6: Qualitative generalization results on LIME and MEF. The examples highlight differences in color fidelity and exposure balance, particularly around the helmet, ceiling details, and bright window regions.
Fig. 7: Qualitative comparison on U3D nighttime aerial scenes. MrLIE reduces artificial-light residues and recovers daytime appearance while retaining the stadium outline and riverbank geometry.
Fig. 9: User study against representative baselines. Each bar reports the percentage of pairwise votes favoring MrLIE over the compared method, indicating stronger perceived naturalness and restoration quality.
Type
Strategy
Metrics
NIQE
BRISQUE
PI
Ref.
One-shot
3.406
11.826
2.324
w/o MLLM-Critic
3.203
12.481
2.210
w/o IQA Scorer
3.249
13.039
2.244
MLLM-Gen
2.957
8.110
2.114
LLIE
One-shot
2.851
12.837
2.322
TABLE IV: Ablation study on MLLM-Gen for reference generation (Ref.) and low-light image enhancement (LLIE). One-shot uses a single generation pass; the partial variants omit MLLM-Critic or the IQA scorer from iterative refinement.
Fig. 10: Visual ablation of MLLM-guided reference generation. The upper row compares generated references and the lower row shows the corresponding restoration results. Red boxes indicate residual artificial-light cues. The complete MLLM-Gen suppresses nighttime-light residues and provides more reliable daytime guidance for restoration.
Var.
Components
Metrics
Conf
Offset
NIQE
BRISQUE
PI
1
×
×
3.234
7.258
2.399
2
×
✓
3.202
4.894
2.306
3
✓
✓
2.884
5.368
2.110
TABLE V: Component ablation of COM. Offset denotes learnable spatial warping of reference features, and Conf denotes confidence-gated fusion for suppressing unreliable reference regions.
Fig. 11: Visualization of COM. For each example, we show the low-light input and feature, the generated reference image, the predicted offset field, the confidence map, and the aligned reference feature.
Fig. 12: Visual ablation of SAI. The variants correspond to Table VI . Removing Attn1 disables the reference interaction and causes severe structural drift, while removing Attn2 or Add weakens the anchoring to the low-light input. The full model preserves the input layout while recovering more reliable structures.
Var.
Components
Metrics
Attn1
Attn2
Add
NIQE
BRISQUE
PI
A
×
✓
✓
3.017
10.196
2.281
B
✓
×
✓
3.073
7.770
2.311
C
✓
✓
×
3.097
5.398
2.233
D
✓
✓
✓
2.884
5.368
2.110
TABLE VI: Component ablation of SAI. Attn1 introduces aligned reference cues, Attn2 re-anchors the reference-enhanced representation to the low-light structure, and Add denotes the ControlNet-style additive branch.
Weights
Metrics
λoff
λconf
NIQE
BRISQUE
PI
0
0
2.995
5.483
2.168
0.001
0
3.069
6.815
2.373
0.001
0.0001
3.204
4.606
2.275
0.001
0.0005
2.884
5.368
2.110
0.001
0.001
3.244
5.066
2.322
TABLE VII: Sensitivity analysis of the COM regularization weights. The two blocks sweep the confidence and offset regularizers around the default setting.
Fig. 13: Iterative refinement of a daytime reference through diagnosis-guided prompt updates. Each candidate is accompanied by diagnostic scores and the resulting action, illustrating how artificial-light suppression and dark-region brightening lead to the selected reference.
Self-supervised low-light image enhancement (LLIE) is highly appealing as it eliminates the reliance on external paired data. However, the lack of external references causes networks to struggle with decoupling entangled illumination, delicate textures, and amplified noise. To resolve this challenge, we propose an Internally Referenced LLIE framework that extracts reliable physical and structural references from the degraded input image itself. First, we introduce a local exposure-simulated scheme to extract a low-frequency pseudo ground-truth. This serves as an internal physical reference to guide global illumination estimation and correct color casts. Second, we propose a dual-domain preservation strategy with spatial and spectral constraints to construct internal structural references. Specifically, an Illumination-Aligned Perceptual loss preserves global structures under illumination shifts, while a Shift-Invariant Spectral Correlation loss captures fine-grained local structures and suppresses high-frequency noise. Finally, we propose a Gain-Adaptive Feature Modulation (GAFM) mechanism to address highly spatially-variant residual noise. By transforming the self-estimated illumination map into an internal spatial gain prior, GAFM dynamically guides a blind-spot network for spatially-aware denoising. Extensive experiments demonstrate that our method achieves state-of-the-art performance, delivering superior noise suppression and textural fidelity. Code will be publicly released at https://visonj.github.io/IRLE/.
Peiyuan He, Hainuo Wang, Hengxing Liu +2
College of Intelligence and Computing, Tianjin University, Tianjin 300350, China
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) aims to recover well-exposed and perceptually faithful images from severely degraded observations. Recent diffusion-based generative methods achieve strong perceptual quality by leveraging powerful image priors, but their iterative sampling incurs substantial computational cost, limiting practical deployment. Despite recent efficiency gains in generative LLIE, high-quality enhancement with a single function evaluation and low overall computation remains challenging, particularly under diverse real-world low-light conditions. We introduce GLOW-FM, a one-step latent flow-matching framework for efficient generative LLIE. GLOW-FM performs conditional transport in the latent space of a frozen Tiny-VAE and leverages Terminal Velocity Matching (TVM) to enable enhancement with a single function evaluation. Matched comparisons with alternative one-step objectives support TVM as an effective design choice, and, to the best of our knowledge, this is its first adaptation to image restoration and specifically to LLIE. Extensive experiments on challenging real-world benchmarks demonstrate that GLOW-FM achieves competitive image quality and state-of-the-art cross-dataset generalization while requiring only a single function evaluation and 34 GFLOPs for the small variant. These findings highlight the potential of GLOW-FM to bring generative LLIE closer to practical deployment through single-step inference and compact latent modeling. Code will be made publicly available upon publication.