Generative strategy for low and normal light image pairs with enhanced statistical fidelity

Abstract

Low light image enhancement remains challenging due to limited availability of real low/normal light image pairs for training. Simple image simulation techniques used for data augmentation fail to accurately model noise and distortions present in real low light photos. In this work, we propose N2LDiff, a novel generative model leveraging diffusion processes to synthesize realistic low light images from normal light counterparts. Our model leverages the noise modeling capabilities of diffusion processes to generate low light images with accurate noise, blurring, and color distortions. We make the following key contributions: (1) We develop a novel N2LDiff model that can generate varied low light images from the same normal light input via diffusion processes. (2) We introduce a new benchmark for low light image synthesis using existing datasets. (3) Leveraging N2LDiff, we construct a large-scale low light dataset. Our generated data will facilitate training and evaluation of deep learning models for low light enhancement tasks.