LightQANet: Quantized and Adaptive Feature Learning for Low-Light Image Enhancement

TL;DR

This paper tackles low-light image enhancement by addressing unreliable feature representations caused by severe illumination losses. It introduces LightQANet, which combines a Light Quantization Module (LQM) that quantizes illumination factors and a Light-Aware Prompt Module (LAPM) that dynamically guides feature learning with illumination priors, built atop a high-quality codebook via VQ-GAN. Key contributions include the light-factor space learning with a light consistency loss, region-aware prompt fusion, and extensive cross-dataset evaluation showing state-of-the-art results. The approach improves robustness and texture/color fidelity across diverse lighting conditions with a favorable complexity/accuracy trade-off.

Abstract

Low-light image enhancement (LLIE) aims to improve illumination while preserving high-quality color and texture. However, existing methods often fail to extract reliable feature representations due to severely degraded pixel-level information under low-light conditions, resulting in poor texture restoration, color inconsistency, and artifact. To address these challenges, we propose LightQANet, a novel framework that introduces quantized and adaptive feature learning for low-light enhancement, aiming to achieve consistent and robust image quality across diverse lighting conditions. From the static modeling perspective, we design a Light Quantization Module (LQM) to explicitly extract and quantify illumination-related factors from image features. By enforcing structured light factor learning, LQM enhances the extraction of light-invariant representations and mitigates feature inconsistency across varying illumination levels. From the dynamic adaptation perspective, we introduce a Light-Aware Prompt Module (LAPM), which encodes illumination priors into learnable prompts to dynamically guide the feature learning process. LAPM enables the model to flexibly adapt to complex and continuously changing lighting conditions, further improving image enhancement. Extensive experiments on multiple low-light datasets demonstrate that our method achieves state-of-the-art performance, delivering superior qualitative and quantitative results across various challenging lighting scenarios.

Figures (14)

• Figure 1: Visual comparisons on the real low-light image. The input image is compared with results from SNRSNR, SMGSMG, RetinexFormerRetinexformer, CIDNetCIDNet, and our method. The zoomed-in regions show the details of texture and sharpness restoration. Our method produces the most natural textures and smooth transitions around edges.
• Figure 2: Effectiveness analysis and benchmark comparison. (a) represents results of the baseline built by VQ-GAN VQGAN and our method. (b) shows code activation frequency on LOL-v2 Real dataset LOLv2. “GT in Baseline” and “LL in Baseline” represent inputting ground truth and low-light images to pre-trained VQ-GAN VQGAN and fine-tuned VQ-GAN with low-light images, respectively. “LL in CodeEnhance” and “LL in Ours” denote inputting low-light images to CodeEnhance xu24 and our method. Activation frequency of “LL in Ours” is closer to that of “GT in Baseline”, which indicates our method can learn and activate important features under low-light conditions. (c) and (d) present performance comparison on LLIE datasets (LOL-v2-Synthetic LOLv2, LOL-v2-Real LOLv2, LSRW-Huawei LSRW, LIME LIME, MEF MEF, VV \ref{['vv']}, and NPE NPE) in terms of PSNR (the higher the better) and NIQE (the lower the better). As highlighted by the bold red line, the proposed method consistently achieves the best demonstrating its superior.
• Figure 3: Overview of our proposed LightQANet framework. Our method leverages a pretrained codebook $\mathbf{C}^{\dag}$ and a decoder ${\rm D}^{\dag}$ as the foundations. In LightQANet, the Light Quantization Module (LQM) is utilized to extract light factors and promote the learning of light-invariant features by a novel light consistency loss ($\mathcal{L}_{lcl}$). Additionally, the Light-Aware Prompt Module (LAPM) is introduced to encode light illumination data for dynamically guiding the feature learning process. Finally, the feautre fusion via linear interpolation refines the reconstructed features.
• Figure 4: Overview of the proposed LQM. (a) and (b) illustrate the alternating optimization of the LQM and the LightQANet. (c) shows the Gram matrices from low-light (LL) and normal-light (NL) images, which exhibit significant overlap. In contrast, the light factors are clearly distinguished based on lighting conditions, underscoring their effectiveness in accurately capturing and representing light-relevant information. Note that the LQM is only used during the training phase and therefore does not impact the processing speed during the inference stage.
• Figure 5: Prompt-Brightness correlation analysis. (a) Correlation heatmap between prompt weights (Prompt Index 1–5) and brightness levels. Positive (red) and negative (blue) values indicate correlation strength. Different prompts show distinct sensitivities, with some strongly responding to extremely dark regions [0, 25.5) and others to brighter or transitional regions, highlighting their complementary roles. (b) Brightness distribution histogram of the MEF dataset MEF, showing a predominance of darker pixels.