Visual Thinking of Neural Networks: Interactive Text to Image Synthesis

Abstract

Reasoning, a trait of cognitive intelligence, is regarded as a crucial ability that distinguishes humans from other species. However, neural networks now pose a challenge to this human ability. Text-to-image synthesis is a class of vision and linguistics, wherein the goal is to learn multimodal representations between the image and text features. Hence, it requires a high-level reasoning ability that understands the relationships between objects in the given text and generates high-quality images based on the understanding. Text-to-image translation can be termed as the visual thinking of neural networks. In this study, our model infers the complicated relationships between objects in the given text and generates the final image by leveraging the previous history. We define diverse novel adversarial loss functions and finally demonstrate the best one that elevates the reasoning ability of the text-to-image synthesis. Remarkably, most of our models possess their own reasoning ability. Quantitative and qualitative comparisons with several methods demonstrate the superiority of our approach.

Keywords

• Generative adversarial networks
• image generation
• multimodal learning
• multimodal representation
• text-to-image synthesis