The paper introduces Masked Autoregressive Flow (MAF), a novel approach to density estimation that leverages autoregressive models and normalizing flows. MAF enhances the flexibility of autoregressive models by stacking multiple instances of the same model, each modeling the random numbers generated by the next model in the stack. This construction results in a normalizing flow that is more flexible than individual autoregressive models while maintaining tractable Jacobians. MAF is closely related to Inverse Autoregressive Flow (IAF) and is a generalization of Real NVP. The authors demonstrate that MAF achieves state-of-the-art performance in various density estimation tasks, outperforming both IAF and Real NVP. The paper also discusses the theoretical connections between MAF, IAF, and Real NVP, and provides experimental results on several datasets, including UCI datasets, image patches, MNIST, and CIFAR-10.The paper introduces Masked Autoregressive Flow (MAF), a novel approach to density estimation that leverages autoregressive models and normalizing flows. MAF enhances the flexibility of autoregressive models by stacking multiple instances of the same model, each modeling the random numbers generated by the next model in the stack. This construction results in a normalizing flow that is more flexible than individual autoregressive models while maintaining tractable Jacobians. MAF is closely related to Inverse Autoregressive Flow (IAF) and is a generalization of Real NVP. The authors demonstrate that MAF achieves state-of-the-art performance in various density estimation tasks, outperforming both IAF and Real NVP. The paper also discusses the theoretical connections between MAF, IAF, and Real NVP, and provides experimental results on several datasets, including UCI datasets, image patches, MNIST, and CIFAR-10.