DiffLIE: Low-Light Image Enhancment based on Deep Diffusion Model

Low-light Image Enhancement (LIE) has received significant attention in the field of computer vision low-level tasks in recent years. Previous work mainly treated LIE as a lighting enhancement work based on the Retinex theory. In this paper, we propose a new approach that utilizes the powerful generative network, the deep diffusion model, to regard LIE as a task of generating normal illumination images from pure black images. We modeled the enhancement process of low-light images based on the Markov assumption of conditional probability and constructed a controllable illumination enhancement model with diversified outputs. The experimental results achieved excellent visual effects and PSNR indicators on various synthetic datasets and real datasets.