Erasing text from images is a common image-editing task in film industry and shared media. Existing text-erasing models either tend to produce artifacts or fail to remove all the text in real-world images. In this paper, we follow a two-stage text erasing framework that first masks the text by segmentation, and then inpaints the masked region to create a text-erased image. Our proposed text mask generator is designed to accurately cover text, which combined with inpainting, can produce reliable text-erased results. In the inpainting model, we propose a Multiscale Gradient Reconstruction Loss to generate sharp realistic-looking images. Our model achieves stateof-the-art results on both synthetic and real world data in both quantitative and qualitative measures.