Nouveau VAE (NVAE) is a deep hierarchical Variational Autoencoder (VAE) designed for image generation. It uses depth-wise separable convolutions and batch normalization to enhance the receptive field and stabilize training. NVAE employs a residual parameterization of Normal distributions and spectral regularization to improve the optimization of the KL term. The model achieves state-of-the-art results on various image datasets, including MNIST, CIFAR-10, CelebA 64, and CelebA HQ, and provides a strong baseline on FFHQ. NVAE is the first VAE to produce high-quality images as large as 256x256 pixels. The paper also discusses the importance of careful neural architecture design in VAEs and presents practical solutions to reduce memory usage and improve training stability.Nouveau VAE (NVAE) is a deep hierarchical Variational Autoencoder (VAE) designed for image generation. It uses depth-wise separable convolutions and batch normalization to enhance the receptive field and stabilize training. NVAE employs a residual parameterization of Normal distributions and spectral regularization to improve the optimization of the KL term. The model achieves state-of-the-art results on various image datasets, including MNIST, CIFAR-10, CelebA 64, and CelebA HQ, and provides a strong baseline on FFHQ. NVAE is the first VAE to produce high-quality images as large as 256x256 pixels. The paper also discusses the importance of careful neural architecture design in VAEs and presents practical solutions to reduce memory usage and improve training stability.