Enhancing Low-Light Image Reconstruction via Non-Autoregressive Transformers: A Mask-Aware Latent Integration Framework

Reconstructing high-quality images under low-light conditions remains a persistent challenge in computer vision due to degradation in contrast, color, and structural information. Existing enhancement techniques, such as Retinex-based decomposition and adversarial learning, often struggle with preserving global semantics and fine-grained perceptual fidelity. This paper introduces a novel dual-path enhancement framework that integrates Non-Autoregressive Transformers (NAT) into two critical stages of the GLARE model: the latent representation alignment and the decoder-level feature transformation. A maskaware mechanism is applied at both stages to identify uncertain or degraded regions, which are iteratively refined using NAT's parallel decoding and re-masking strategy. This dual integration enables the model to capture global context, improve detail recovery, and enhance structural consistency without altering the original architecture. Experiments conducted on benchmark datasets, including GLARE and LOL, demonstrate that the proposed method achieves superior performance, with PSNR improving from 25.52 to 27.08, SSIM increasing from 0.8462 to 0.8678, and LPIPS decreasing from 0.1299 to 0.1162. These improvements validate the effectiveness of NAT-based refinement at multiple levels and highlight the potential of integrating codebook-guided generation with transformer-based reasoning in real-world low-light environments, such as industrial inspection, automated surveillance, and smart manufacturing systems.