In recent years, Generative Adversarial Network has rapidly developed in the field of image generation and image inpainting. In this project, we explore a new deep learning-based inpainting method, which can not only generate new image structure through convolution but can also borrow patches from the surrounding region. The report will analyze the architecture and structures of the model and experiments with the model on two datasets (Chart and diagram dataset, 2D molecule structure dataset). The result demonstrates that the model generated relatively high-quality inpainting images for both datasets, although there are still blurry result and chromatic aberration in some regions.