In deep learning, generative models is a rapidly growing research field, where GAN (generative adversarial nets) and VAE (variational autoencoder) are two of the most popular approaches. Given a data set, GAN models the probability density only implicitly, and VAE involves approximation from dimensional reduction. Learning the explicit probability density function is a hard task. In this tutorial, we will talk about how to tackle this problem by deep normalizing flows, with applications to outlier detection by density estimation, data characterization by latent variable inference, and completing the missing features such as image inpainting.