This paper presents a deep neural network called KinD for low-light image enhancement. The network is inspired by Retinex theory and decomposes images into two components: illumination and reflectance. This decomposition allows the original image space to be separated into two smaller subspaces, which can be more effectively regularized and learned. The network is trained using paired images captured under different light conditions, without relying on ground-truth illumination or reflectance information. The design enables flexible adjustment of light levels and effectively removes degradations such as noise and color distortion. The network is efficient, processing images in VGA resolution in less than 50ms on a 2080Ti GPU. Extensive experiments demonstrate that KinD outperforms state-of-the-art methods in terms of image quality metrics like PSNR, SSIM, and NIQE. The model is robust against severe visual defects and user-friendly for adjusting light levels. The network also addresses the challenge of handling varying noise levels in different regions of an image, and it effectively removes color distortion. The proposed method provides a flexible mapping function for adjusting light levels, offering users greater control over the enhancement process. The results show that KinD produces visually pleasing images with improved contrast and reduced noise, making it suitable for practical applications.This paper presents a deep neural network called KinD for low-light image enhancement. The network is inspired by Retinex theory and decomposes images into two components: illumination and reflectance. This decomposition allows the original image space to be separated into two smaller subspaces, which can be more effectively regularized and learned. The network is trained using paired images captured under different light conditions, without relying on ground-truth illumination or reflectance information. The design enables flexible adjustment of light levels and effectively removes degradations such as noise and color distortion. The network is efficient, processing images in VGA resolution in less than 50ms on a 2080Ti GPU. Extensive experiments demonstrate that KinD outperforms state-of-the-art methods in terms of image quality metrics like PSNR, SSIM, and NIQE. The model is robust against severe visual defects and user-friendly for adjusting light levels. The network also addresses the challenge of handling varying noise levels in different regions of an image, and it effectively removes color distortion. The proposed method provides a flexible mapping function for adjusting light levels, offering users greater control over the enhancement process. The results show that KinD produces visually pleasing images with improved contrast and reduced noise, making it suitable for practical applications.