Illuminating Darkness: Learning to Enhance Low-light Images In-the-Wild

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

Abstract

Single-shot low-light image enhancement (SLLIE) remains challenging due to the limited availability of diverse, real-world paired datasets. To bridge this gap, we introduce the Low-Light Smartphone Dataset (LSD), a large-scale, high-resolution (4K+) dataset collected in the wild across a wide range of challenging lighting conditions (0.1 to 200 lux). LSD contains 6,425 precisely aligned low and normal-light image pairs, selected from over 8,000 dynamic indoor and outdoor scenes through multi-frame acquisition and expert evaluation. To evaluate generalization and aesthetic quality, we collect 2,117 unpaired low-light images from previously unseen devices. To fully exploit LSD, we propose TFFormer, a hybrid model that encodes luminance and chrominance (LC) separately to reduce color-structure entanglement. We further propose a cross-attention-driven joint decoder for context-aware fusion of LC representations, along with LC refinement and LC-guided supervision to significantly enhance perceptual fidelity and structural consistency. TFFormer achieves state-of-the-art results on LSD (+2.45 dB PSNR) and substantially improves downstream vision tasks, such as low-light object detection (+6.80 mAP on ExDark).

The Challenge of Real-World Low-Light Enhancement

Single-shot low-light image enhancement (SLLIE) is often considered the "holy grail" of smartphone photography. When capturing images in dark environments, sensors struggle with a low signal-to-noise ratio, leading to significant noise and color distortion. Current AI models frequently fail in real-world "in-the-wild" scenarios because they are typically trained on small, synthetic, or highly controlled datasets.

Existing datasets often rely on artificially dimmed images or low-ISO settings that do not accurately reflect real-world sensor noise. Consequently, models trained on such data often produce over-brightened areas, unrealistic color shifts, or amplified noise when applied to actual low-light photos. The authors identify two primary gaps: a lack of diverse, high-quality training data and model architectures that "entangle" color and structure, leading to blurry textures and unnatural hues.

The Low-Light Smartphone Dataset (LSD)

To address the data scarcity, the authors developed the Low-Light Smartphone Dataset (LSD). This is currently the largest real-world dataset for low-light enhancement, featuring 6,425 paired 4K+ images captured across 0.1 to 200 lux.

Using a custom Android application, the team captured precisely aligned image pairs using 15 different smartphone sensors. They employed a "three-frame capture" strategy to ensure alignment and a VGG-based scene classifier to automatically filter out low-quality training pairs (e.g., those with motion blur or defocus). This curated approach proved vital; training on the filtered dataset led to a +2.14 dB PSNR improvement over training on raw, unfiltered data.

TFFormer: Decoupling Luminance and Chrominance

To fully leverage the LSD dataset, the authors propose TFFormer, a hybrid Transformer-CNN architecture. The core innovation of TFFormer is its "tuning-fork" shape, which decouples the processing of brightness and color to prevent "color-structure entanglement."

The TFFormer architecture features decoupled luminance and chrominance branches to prevent color-structure entanglement.

In standard RGB models, noise in one color channel can bleed into the structural representation of an object. TFFormer avoids this by separating the image into Luminance (L) and Chrominance (C) streams:

  • Decoupled Encoding: Dedicated encoders process L (focusing on structure and denoising) and C (focusing on color fidelity) separately.
  • LCCAB (Cross-Attention Block): A novel fusion module uses cross-attention to let the luminance features guide the chrominance features, ensuring colors align perfectly with structural edges.
  • LCGRB (Refinement Block): A final stage uses LC-aware attention to polish the final output.
  • LCCG Loss: A specific loss function penalizes errors in luminance and chrominance independently, preventing the over-smoothing common in traditional models.

Experimental Results and Analysis

TFFormer achieves state-of-the-art results while maintaining high efficiency. With only 5.87M parameters, it is significantly smaller than heavy-duty models like MIRNet (which is 5.4x larger and 24x slower) yet performs better on the LSD benchmark. Quantitatively, TFFormer outperformed the previous best model, Retinexformer, by +2.04 dB PSNR and +0.0738 SSIM on the LSD dataset.

Qualitative comparison of TFFormer against existing low-light enhancement methods on the LSD dataset.

Beyond standard metrics, the model's performance translates to real-world utility. Enhancing images with LSD-trained models boosted object detection performance (YOLOv3) on the ExDark dataset by +6.80 mAP. Furthermore, in a blind user study, participants preferred TFFormer’s results over 82% of the time, citing more natural colors and sharper structural details.

Conclusion

While TFFormer represents a significant advancement, challenges remain in extreme low-light conditions (under 1 lux), where visible noise or "halos" can still occur. Future research will focus on extending the LSD dataset for more complex downstream tasks like semantic segmentation in the dark and refining performance in near-total darkness through joint denoising and enhancement strategies. By providing both a massive, high-quality dataset and a novel decoupled architecture, this work sets a new benchmark for real-world low-light image enhancement.

Summary updated

Abstract

Single-shot low-light image enhancement (SLLIE) remains challenging due to the limited availability of diverse, real-world paired datasets. To bridge this gap, we introduce the Low-Light Smartphone Dataset (LSD), a large-scale, high-resolution (4K+) dataset collected in the wild across a wide range of challenging lighting conditions (0.1 to 200 lux). LSD contains 6,425 precisely aligned low and normal-light image pairs, selected from over 8,000 dynamic indoor and outdoor scenes through multi-frame acquisition and expert evaluation. To evaluate generalization and aesthetic quality, we collect 2,117 unpaired low-light images from previously unseen devices. To fully exploit LSD, we propose TFFormer, a hybrid model that encodes luminance and chrominance (LC) separately to reduce color-structure entanglement. We further propose a cross-attention-driven joint decoder for context-aware fusion of LC representations, along with LC refinement and LC-guided supervision to significantly enhance perceptual fidelity and structural consistency. TFFormer achieves state-of-the-art results on LSD (+2.45 dB PSNR) and substantially improves downstream vision tasks, such as low-light object detection (+6.80 mAP on ExDark).
BibTeX
@inproceedings{sharif2026illuminating,
  author = {S M A Sharif and Abdur Rehman and Zain Ul Abidin and Fayaz Ali Dharejo and Radu Timofte and Rizwan Ali Naqvi},
  title = {{Illuminating Darkness: Learning to Enhance Low-light Images In-the-Wild}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2026},
}
Figure