Enriching Attributes from Knowledge Graph for Fine-grained Text-to-Image Synthesis

Abstract

In this paper, we propose an Attribute-Rich Generative Adversarial Network (AttRiGAN) for text-to-image synthesis, which enriches the simple text description by associating knowledge graph and embedding it in the synthesis task in the form of an attribute matrix. Higher fine-grained images can be synthesized with AttRiGAN, and the synthesized sample are more similar to the objects that exist in the real world, since they are driven by attributes which are enriched from the knowledge graph. The experiments conducted on two widely-used fine-grained image datasets show that our AttRiGAN allows a significant improvement in fine-grained text-to-image synthesis.