The paper proposes an AI model to generate accurate normal maps for portrait images. We utilized a portrait photo booth system based on photometric methods to generate training data. With these data, users only need to input a portrait image, then it can be converted into normal map image which is further used to relight the color portrait image.