Restoring images from low-light data is a challenging problem. Most existing deep-network based algorithms are designed to be trained with pairwise images. Due to the lack of real-world datasets, they usually perform poorly when generalized in practice in terms of loss of image edge and color information. In this paper, we propose an unsupervised generation network with attention-guidance to handle the low-light image enhancement task. Specifically, our network contains two parts: an edge auxiliary module that restores sharper edges and an attention guidance module that recovers more realistic colors. Moreover, we propose a novel loss function to make the edges of the generated images more visible. Experiments validate that our proposed algorithm performs favorably against state-of-the-art methods, especially for real-world images in terms of image clarity and noise control.