Low-light image enhancement with diagonal frequency feature refinement and truncated sampling in conditional diffusion models

Low-light image enhancement (LLIE) aims to improve the visibility and contrast of images captured under insufficient illumination conditions, which is critical for various applications such as surveillance, autonomous driving, and medical imaging. Recently, diffusion models have shown strong generative capabilities and have emerged as a promising potential in the field of LLIE. However, existing diffusion-based methods for LLIE still suffer from several critical challenges: (1) The iterative denoising process inherent to diffusion models leads to high computational cost, limiting their applicability in real-time scenarios; (2) Spatial-domain diffusion methods struggle to recover the high-frequency textures and structural details in severely degraded low-light images, resulting in enhanced results with blurriness or visual artifacts; (3) The accumulation of errors during the reverse diffusion process introduces color shifts and inconsistencies, affecting the naturalness and perceptual quality of the enhanced images. To address these challenges, this paper proposes a low-light image enhancement with diagonal frequency feature refinement and truncated sampling in conditional diffusion models, which aims to improve enhancement quality while significantly reducing computational overhead. Specifically, the proposed method conducts the diffusion process in the low-frequency subspace of the wavelet domain, which not only preserves the global structural and illumination information of the image but also reduces spatial dimensionality and computational complexity. To compensate for the directional sparsity and loss of fine details in the high-frequency components, a high-frequency refinement module is introduced. This module leverages the horizontal and vertical high-frequency features to reconstruct the missing diagonal textures via a dual-path cross-directional attention mechanism, thereby enhancing local texture fidelity and detail preservation. Furthermore, an efficient truncated sampling strategy is employed to accelerate the reverse diffusion process. By adopting a step-interrupt mechanism, our method substantially reduces inference time and memory consumption without degrading visual quality. To further enhance stability and color fidelity in the enhanced results, a contrast-aware correction module is integrated into the denoising stages of reverse diffusion. By utilizing contrast and color priors from previous steps, this module adaptively adjusts the color and brightness at each sampling step.Result To validate the effectiveness of the proposed method, extensive experiments are conducted on three publicly available paired low-light datasets, comparing the proposed method with several state-of-the-art LLIE methods. Both subjective visual comparisons and objective quantitative metrics demonstrate that the proposed method consistently outperforms existing approaches. On the LOLv2-Real dataset, the proposed method achieves improvements over the second-best method by 8.46 % in peak signal-to-noise ratio (PSNR), 1.22% in structural similarity index (SSIM), a 10.19% reduction in learned perceptual image patch similarity (LPIPS), and a 0.38% reduction in naturalness image quality evaluator (NIQE), indicating better perceptual quality and structural fidelity. To assess the cross-domain generalization ability, we further evaluated the model using the pretrained weights from LOLv2-Real datasets on four unpaired low-light datasets. The proposed method achieved lower NIQE scores than existing methods by 0.82 (low-light image enhancement via illumination map estimation, LIME), 0.37 (digital images from commercial cameras, DICM), 0.39(multi-exposure image fusion, MEF) and 0.86 (naturalness photo enhancement, NPE), demonstrating strong generalization ability without target-domain fine-tuning. Additionally, ablation studies confirm the contribution of each core module. Compared with spatial-domain diffusion baselines, the proposed method significantly reduced inference time from 7.617 seconds to 0.473 seconds per image, and memory usage decreased by 48.53 %, highlighting the efficiency and practicality of the proposed design. The proposed method integrates frequency-domain modeling with an efficient diffusion sampling strategy, introducing three key innovations. (1) By performing the diffusion process in the low-frequency subspace of the wavelet domain, which significantly reduces computational complexity while preserving global structural and illumination information. (2) A dedicated high-frequency refinement module is developed to recover diagonal texture details by leveraging horizontal and vertical high-frequency information, thereby improving the sharpness and perceptual quality of the enhanced images. (3) The integration of an efficient truncated sampling strategy and a contrast-aware correction module improves sampling efficiency and color fidelity, effectively balancing enhancement quality and computational cost. Collectively, these contributions enable the proposed method to achieve superior enhancement performance with substantially reduced inference overhead, demonstrating strong potential for real-world low-light vision applications.