Masked Autoregressive Flow (MAF) is a type of normalizing flow that improves the flexibility of autoregressive models by stacking multiple models to model the random numbers used internally during data generation. This approach is closely related to Inverse Autoregressive Flow (IAF) and generalizes Real NVP. MAF achieves state-of-the-art performance in density estimation tasks by using the Masked Autoencoder for Distribution Estimation (MADE) as a building block, enabling fast evaluation and training on parallel architectures like GPUs. MAF is more suitable for density estimation than IAF, as it requires only one pass through the model for density evaluation, whereas IAF requires multiple passes. MAF can also be extended to conditional density estimation by incorporating additional input variables. Experiments show that MAF outperforms Real NVP and achieves state-of-the-art results on various density estimation tasks, including UCI datasets and image patches. MAF is a flexible and efficient method for density estimation, with applications in probabilistic programming, variational inference, and likelihood-free inference.Masked Autoregressive Flow (MAF) is a type of normalizing flow that improves the flexibility of autoregressive models by stacking multiple models to model the random numbers used internally during data generation. This approach is closely related to Inverse Autoregressive Flow (IAF) and generalizes Real NVP. MAF achieves state-of-the-art performance in density estimation tasks by using the Masked Autoencoder for Distribution Estimation (MADE) as a building block, enabling fast evaluation and training on parallel architectures like GPUs. MAF is more suitable for density estimation than IAF, as it requires only one pass through the model for density evaluation, whereas IAF requires multiple passes. MAF can also be extended to conditional density estimation by incorporating additional input variables. Experiments show that MAF outperforms Real NVP and achieves state-of-the-art results on various density estimation tasks, including UCI datasets and image patches. MAF is a flexible and efficient method for density estimation, with applications in probabilistic programming, variational inference, and likelihood-free inference.