Controllable Text-to-Image Generation with Enhanced Text Encoder and Edge-Preserving Embedding

Abstract

Text-to-image generation through generative adversarial networks (GANs) has become a popular research topic in the fields of natural language processing and high-quality image synthesizing. By learning latent word features of queries, many state-of-the-art GANs generate images from coarse to fine in a multi-stage trend. While this mechanism is able to synthesize realistic images stage by stage, failures containing semantic mismatches and shape distortions still exist. In this project, we adopted a multi-stage text-to-image GAN as a baseline, and remodeled this structure in two ways: 1) modify the text encoder by pre-processing text queries with pre-trained language models to gain deeply mined text information, and 2) develop an edge-preserving structure to capture feature mismatches, denoted as “edge loss”, between text latent information and generated image edges. Experiments on benchmark dataset demonstrate that models of our method outperform the baseline model and increase Inception Scores by over 30%, and that our approaches are able to effectively generate synthetic images using natural language descriptions.