Organizations: Xi’an University of Electronic Science and Technology, China · National University of Defense Technology, China · School of Computer Science and Engineering, Nanjing University of Science and Technology, China · School of Information Science and Engineering, Shandong Normal University, China · School of Computing and Artificial Intelligence, Shandong University of Finance and Economics, China · the College of Information Science and Engineering, Shandong Agricultural University, China · Xi’an University, China · Qilu University of Technology, China
Considering efficiency, ultra-high-definition (UHD) low-light image restoration is extremely challenging. Existing methods based on Transformer architectures or high-dimensional complex convolutional neural networks often suffer from the "memory wall" bottleneck, failing to achieve millisecond-level inference on edge devices. To address this issue, we propose a novel real-time UHD low-light enhancement network based on geometric feature fusion using Clifford algebra in 2D Euclidean space. First, we construct a four-layer feature pyramid with gradually increasing resolution, which decomposes input images into low-frequency and high-frequency structural components via a Gaussian blur kernel, and adopts a lightweight U-Net based on depthwise separable convolution for dual-branch feature extraction. Second, to resolve structural information loss and artifacts from traditional high-low frequency feature fusion, we introduce spatially aware Clifford algebra, which maps feature tensors to a multivector space (scalars, vectors, bivectors) and uses Clifford similarity to aggregate features while suppressing noise and preserving textures. In the reconstruction stage, the network outputs adaptive Gamma and Gain maps, which perform physically constrained non-linear brightness adjustment via Retinex theory. Integrated with FP16 mixed-precision computation and dynamic operator fusion, our method achieves millisecond-level inference for 4K/8K images on a single consumer-grade device, while outperforming state-of-the-art (SOTA) models on several restoration metrics.
Figures & tables
Figure 1. Comprehensive performance comparison of low-light enhancement models at 4K resolution. The x-axis and bubble size indicate inference latency and peak VRAM consumption (in log scale), respectively. The y-axis represents the reconstruction quality (PSNR). The green shaded area highlights the real-time processing regime ( \textgreater 30 FPS). Similar performance trends at 8K resolution are provided in the Appendix. A scatter plot comparing inference speed, PSNR, and memory usage of various models at 4K resolution.
Figure 2. Overview of the proposed Clifford Pyramid Enhance (CPE) framework. (a) The overall pipeline based on a multi-scale pyramid. (b1) The frequency-decoupled feature extraction module (Pyramid Layer) for separating low-frequency illumination and high-frequency structural features. (b2) Details of the dual-branch LightUNet.
Figure 3. Details of the geometric feature fusion module based on Clifford Algebra Cl(2,0).
Figure 4. Dynamic resolution reconstruction and adaptive parameter mapping based on Retinex theory. (a) Original input. (b) Extracted initial illumination map. (c) Predicted Gamma map. (d) Predicted Gain map. (e) Enhanced output. (f) Ground Truth.
Figure 5. Visual comparison between our CPE and other mainstream methods in UHD low-light scenarios. Visual comparison between our CPE and other mainstream methods.
Model
Params (M) ↓
Strategy
4K (3840 × 2160)
8K (7680 × 4320)
Time (ms) ↓
FPS ↑
Time (ms) ↓
FPS ↑
LLFormer ( Wang et al., 2023 )
24.55
Native/Patching
35131.5
0.028
169990.5
0.006
HVI-CIDNet ( Yan et al., 2025 )
1.98
Native/Patching
456.79
2.19
1921.65
0.52
UHDM ( Yu et al., 2022 )
5.93
Native/Resize 256
451.9
2.21
24.13
41.43
Retinexformer ( Cai et al., 2023 )
1.61
Patching
426.9
2.34
3037.6
0.322
FLOL ( Benito et al., 2025 )
0.09
Native
109.0
9.17
421.2
2.37
Table 1. Comparison of computational efficiency (Params, Inference Time, and FPS) in 4K and 8K UHD scenarios. The inference time is measured in milliseconds (ms). The best, second best, and third best results are highlighted in red , blue , and purple , respectively. Notice that heavy models like LLFormer ( Wang et al., 2023 ) require over 160 seconds per 8K image, making them impractical for real-time applications.
Model
Strategy
Metrics
PSNR ↑
SSIM ↑
VIF ↑
LPIPS ↓
NIQE ↓
Panel A: Evaluation at 4K Resolution (3840 × 2160)
LLFormer ( Wang et al., 2023 )
Native
37.31
0.9923
0.9036
0.0194
4.97
HVI-CIDNet ( Yan et al., 2025 )
Native
16.44
0.8776
0.7530
0.1702
6.78
UHDM ( Yu et al., 2022 )
Native
20.12
0.8968
0.5582
0.1428
5.39
Retinexformer ( Cai et al., 2023 )
Native
15.75
0.6851
0.6880
0.1949
4.35
Table 2. Quantitative comparison of low-light enhancement performance on 4K and 8K UHD datasets. ‘Native‘ denotes original resolution input, ‘Patching‘ indicates overlap sliding window strategies, and ‘Resize‘ signifies global downsampling to avoid out-of-memory (OOM). ↑ means higher is better, and ↓ means lower is better. The best, second best, and third best results are highlighted in red , blue , and purple , respectively.
Figure 6. Evaluation of global illumination recovery in extremely dark wide-field scenes.
Figure 7. High-frequency detail reconstruction comparison at native 8K/4K resolution.
Exp.
Proposed Components
PSNR ↑
SSIM ↑
Params (M) ↓
VRAM (MB) ↓
Latency (ms) ↓
FPS ↑
Baseline
Pyramid
Freq-Sep
Illu-Map
Clifford
1
✓
21.56
0.9212
0.434
1143.16
11.36
88.00
2
✓
✓
21.58
0.9313
0.572
1145.34
12.01
83.27
3
✓
✓
21.70
0.9297
0.868
768.29
12.26
81.53
4
✓
✓
✓
21.61
0.9274
0.868
771.60
12.01
83.24
5
✓
✓
✓
20.89
0.9326
1.143
1156.35
13.47
74.25
Table 3. Quantitative results of the ablation study. We progressively integrate the Multi-scale Pyramid (Pyramid), Frequency Decoupling (Freq-Sep), Retinex Illumination Mapping (Illu-Map), and Clifford Geometric Fusion (Clifford) into the baseline network. Evaluation is performed on the 8K UHD dataset.
Figure 8. Demonstration of color constancy and global brightness enhancement by CPE at native 4K and 8K resolutions.
Figure 9. Application evaluation of different enhancement algorithms on the downstream nighttime object detection task (using YOLOv10x). Bounding boxes indicate the detection and confidence scores of key objects such as pedestrians and vehicles.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 10. Comprehensive performance comparison of low-light enhancement models at 8K resolution. The x-axis and bubble size indicate inference latency and peak VRAM consumption (in log scale), respectively. The y-axis represents the reconstruction quality (PSNR). The green shaded area highlights the real-time processing regime ( \textgreater 30 FPS). A scatter plot comparing inference speed, PSNR, and memory usage of various models at 8K resolution.
Figure 11. Additional visual comparison between our CPE and other mainstream methods at native 4K resolution. Best viewed by zooming in.
Figure 12. Hardware overhead evaluation of representative models at native 1080p, 4K, and 8K resolutions. (a) Single-frame inference latency. (b) Peak GPU memory footprint (VRAM). Grey blocks labeled with “OOM” indicate Out-of-Memory failures on a 24GB RTX 3090.
Figure 13. Application evaluation of different enhancement algorithms on the downstream nighttime object detection task at native 4K resolution (using YOLOv10x). Bounding boxes indicate the detection and confidence scores of key objects such as pedestrians and vehicles. This visually complements Figure 9 in the main manuscript, which demonstrates the corresponding detection results at 8K resolution.
Method
Person
Vehicle
Overall
Count ↑
Conf. ↑
Count ↑
Conf. ↑
Count ↑
Conf. ↑
Panel A: Evaluation at 4K Resolution (3840 × 2160)
Input
427
0.7475
1655
0.7803
2082
0.7736
LLFormer ( Wang et al., 2023 )
507
0.7482
1789
0.7965
2296
0.7859
Retinexformer ( Cai et al., 2023 )
431
0.7380
1658
0.7680
2089
0.7618
HVI-CIDNet ( Yan et al., 2025 )
497
0.7486
1748
0.7969
2245
0.7862
Appendix
Table 4. Downstream object detection (YOLOv10x) on UHD-LOL. Red , blue , and purple denote the 1st, 2nd, and 3rd best, respectively.
Figure 14. Mobile Deployment on Huawei Mate 60 Pro
Trial #
Model-only Inference
End-to-End Latency
Panel A: Evaluation on Apple iPhone
1
146
156
2
147
159
3
147
159
4
142
152
5
148
159
Appendix
Table 5. Comparison of Inference Latency on Mobile Devices (ms)
Model
Illumination Naturalness
Seam Artifacts
Detail Blur
Average
Panel A: Evaluation at 4K Resolution ( 3840×2160 )
Ours (CPE)
4.85
4.65
4.85
4.78
LLFormer ( Wang et al., 2023 )
4.85
4.60
4.85
4.77
Retinexformer ( Cai et al., 2023 )
3.50
4.60
2.40
3.50
Zero-DCE++ ( Li et al., 2021 )
2.90
4.60
4.15
3.88
FLOL ( Benito et al., 2025 )
4.50
4.50
4.10
4.37
Appendix
Table 6. Mean Opinion Scores (MOS) at 4K and 8K Resolutions
Shenzhen Campus of Sun Yat-Sen University, China · The Hong Kong Polytechnic University, Hong Kong · Nanjing University of Science and Technology, China +2
State Key Laboratory of Novel Software Technology, Nanjing University, Nanjing 210023, China · School of Computer Science and Technology, Dalian University of Technology, Dalian 116023, China · Department of Computer Science and Engineering, Michigan State University, East Lansing, MI 48824, U.S.A.