Pal-GAN: Palette-conditioned Generative Adversarial Networks

Abstract

Recent advances in Generative Adversarial Networks (GANs) have shown great progress on a large variety of tasks. A common technique used to yield greater diversity of samples is conditioning on class labels. Conditioning on high-dimensional structured or unstructured information has also been shown to improve generation results, e.g. Image-to-Image translation. The conditioning information is provided in the form of human annotations, which can be expensive and difficult to obtain in cases where domain knowledge experts are needed. In this paper, we present an alternative: conditioning on low-dimensional structured information that can be automatically extracted from the input without the need for human annotators. Specifically, we propose a Palette-conditioned Generative Adversarial Network (Pal-GAN), an architecture-agnostic model that conditions on both a colour palette and a segmentation mask for high quality image synthesis. We show improvements on conditional consistency, intersection-over-union, and Fréchet inception distance scores. Additionally, we show that sampling colour palettes significantly changes the style of the generated images.