A Conditional Deep Framework for Automatic Layout Generation

Abstract

Automatic layout generation, which means making computers enjoy creativity, is difficult yet exciting work. Up to now, how to generate reasonable and visually appealing layouts remains a complex challenge. In this paper, we propose a novel layout generation model based on Conditional Generative Adversarial Networks (L-CGAN), which can generate layouts simply and efficiently by positioning, scaling, and flipping the given primitives. To break the bottleneck of limitation of the fixed input size of Generative Adversarial Networks, we develop a pre-processing algorithm to enable the model to generate layouts with an unrestricted number of input elements. Moreover, a graph-constraint module is proposed to guide layout optimization. We demonstrate the competitive performance of our designs in diverse data domains such as handwriting digit layout generation (MNIST Layouts), scene layout generation (AbstractScene-Layouts), and document layout generation (PubLayNet).

Keywords

• Deep learning
• layout generation
• conditional generative adversarial networks
• abstract scene layout