Very Deep VAEs Generalize Autoregressive Models and Can Outperform Them\n on Images

Abstract

We present a hierarchical VAE that, for the first time, generates samples\nquickly while outperforming the PixelCNN in log-likelihood on all natural image\nbenchmarks. We begin by observing that, in theory, VAEs can actually represent\nautoregressive models, as well as faster, better models if they exist, when\nmade sufficiently deep. Despite this, autoregressive models have historically\noutperformed VAEs in log-likelihood. We test if insufficient depth explains why\nby scaling a VAE to greater stochastic depth than previously explored and\nevaluating it CIFAR-10, ImageNet, and FFHQ. In comparison to the PixelCNN,\nthese very deep VAEs achieve higher likelihoods, use fewer parameters, generate\nsamples thousands of times faster, and are more easily applied to\nhigh-resolution images. Qualitative studies suggest this is because the VAE\nlearns efficient hierarchical visual representations. We release our source\ncode and models at https://github.com/openai/vdvae.\n