Glow: Generative Flow with Invertible 1x1 Convolutions

TL;DR

Glow introduces an invertible 1x1 convolution in flow-based generative models, significantly improving log-likelihood performance and enabling efficient, realistic image synthesis and manipulation.

Contribution

It presents a novel invertible 1x1 convolution technique that enhances flow-based generative models' performance and capabilities.

Findings

Significant improvement in log-likelihood on benchmarks

Efficient realistic image synthesis and manipulation

Model is simple and highly parallelizable

Abstract

Flow-based generative models (Dinh et al., 2014) are conceptually attractive due to tractability of the exact log-likelihood, tractability of exact latent-variable inference, and parallelizability of both training and synthesis. In this paper we propose Glow, a simple type of generative flow using an invertible 1x1 convolution. Using our method we demonstrate a significant improvement in log-likelihood on standard benchmarks. Perhaps most strikingly, we demonstrate that a generative model optimized towards the plain log-likelihood objective is capable of efficient realistic-looking synthesis and manipulation of large images. The code for our model is available at https://github.com/openai/glow