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

TL;DR

This paper introduces a very deep hierarchical VAE that surpasses autoregressive models like PixelCNN in likelihood, generates samples faster, and scales effectively to high-resolution images by learning hierarchical representations.

Contribution

It demonstrates that increasing the depth of VAEs enables them to outperform autoregressive models in likelihood and sampling speed on image benchmarks.

Findings

Very deep VAEs outperform PixelCNN in likelihood on CIFAR-10, ImageNet, and FFHQ.

Deep VAEs generate samples thousands of times faster than PixelCNN.

Hierarchical VAEs learn efficient visual representations.

Abstract

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