Illuminating Darkness: Enhancing Real-world Low-light Scenes with Smartphone Images

S. M. A. Sharif, Abdur Rehman, Zain Ul Abidin, Rizwan Ali Naqvi, Fayaz Ali Dharejo, Radu Timofte

Abstract

Low-light image enhancement (LLIE) is a critical component of modern smartphone photography, yet existing deep learning models often fail to generalize to real-world conditions. This paper addresses the "generalization gap" by introducing the Low-Light Smartphone Dataset (LSD), the largest high-quality 4K+ paired dataset to date, and TFFormer, a novel hybrid Transformer-CNN architecture. TFFormer decouples luminance and chrominance processing to prevent color-structure entanglement, a common pitfall in standard RGB-based models. Our approach achieves state-of-the-art results, improving PSNR by over 2 dB and significantly enhancing performance in downstream object detection tasks.

Introduction

Single-shot low-light image enhancement (SLLIE) aims to transform dark, noisy captures into clear, vibrant images. Despite recent progress, many models produce "over-brightened" results, unnatural color shifts, or amplified noise when applied to "in-the-wild" smartphone images. The authors identify two primary causes: the scarcity of diverse, real-world training data and the entanglement of color and structural features in standard processing pipelines.

Existing datasets are often synthetic or captured in controlled environments, failing to reflect the high-ISO noise and spatially varying light typical of mobile photography. Furthermore, processing RGB channels together often leads to a trade-off where fixing structural noise ruins color fidelity.

The Low-Light Smartphone Dataset (LSD)

To bridge the data gap, the authors present the Low-Light Smartphone Dataset (LSD). This massive collection includes 6,425 precisely aligned image pairs at 4K+ resolution, captured across 8,000 scenes using 15 different smartphone sensors. The dataset covers both indoor and outdoor environments with light levels ranging from 0.1 to 200 lux. It includes variants such as DLI (denoised by the phone's ISP) and NLI (raw, noisy captures) to help models handle different levels of degradation.

A key innovation in the dataset's creation was the use of a VGG-based scene classifier to filter out low-quality training patches (e.g., those with motion blur or extreme darkness). This refinement process removed 14% of the raw data but led to a significant boost in final model performance, proving that data quality is as vital as quantity.

TFFormer: Decoupled Luminance and Chrominance

The proposed TFFormer architecture utilizes a "tuning-fork" design to handle light and color separately. By decoupling these components, the model can suppress noise in the structural domain without compromising color accuracy.

The TFFormer architecture decouples luminance and chrominance through separate encoding branches and a cross-attention decoding stage.

The architecture features:

  • LC Decoupling: Inputs are split into a Luminance (L) map for structure and a Chrominance (C) prior for color.
  • LC Guided Attention (LCGAB): Uses semantic cues to guide feature extraction, preserving fine structural details.
  • LCCAB (Cross-Attention): Adaptively fuses features from both branches during decoding to ensure perceptual coherence.
  • LC Consistency Guidance (LCCG): A specialized loss function that prevents the over-smoothing common in traditional L1/L2 losses.

Experimental Results

TFFormer sets a new benchmark for low-light enhancement. Quantitatively, it outperforms the previous state-of-the-art (Retinexformer) by +2.04 dB PSNR and +0.0738 SSIM on the LSD dataset. Qualitatively, TFFormer produces images with more natural colors and sharper details compared to models trained on older datasets like LOL or NCLLIE.

Qualitative comparison showing how the proposed approach (rightmost column) avoids the artifacts and color shifts seen in other methods.

Beyond aesthetics, the enhancement significantly benefits downstream tasks. Using TFFormer as a pre-processor for YOLOv3 on the ExDark dataset resulted in a +6.80 mAP gain. Remarkably, the model achieves these results with a lightweight footprint of only 5.87M parameters and 34.6 GFLOPs, making it significantly more efficient than heavyweights like MIRNet.

Conclusion

By combining a high-quality, diverse dataset with an architecture that respects the physical separation of light and color, the authors provide a robust solution for real-world low-light enhancement. While challenges remain in extreme low-light conditions (< 1 lux), TFFormer and the LSD dataset represent a major step toward laboratory-grade photography in the palm of your hand.

Summary updated

BibTeX
@inproceedings{sharif2026illuminating,
  author = {S. M. A. Sharif and Abdur Rehman and Zain Ul Abidin and Rizwan Ali Naqvi and Fayaz Ali Dharejo and Radu Timofte},
  title = {{Illuminating Darkness: Enhancing Real-world Low-light Scenes with Smartphone Images}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2026},
}
Figure