Realistic Image Generation from Text by Using BERT-Based Embedding

Recently, in the field of artificial intelligence, multimodal learning has received a lot attention due to expectations for enhancement AI performance and potential applications. Text-to-image generation, which is one tasks, challenging topic computer vision natural language processing. The text-to-image generation model based on generative adversarial network (GAN) utilizes text encoder pre-trained with image-text pairs. However, encoders pairs cannot obtain rich information about texts not seen during pre-training, thus it hard generate an image that semantically matches given description. In this paper, we propose new using BERT, widely used BERT as by performing fine-tuning large amount text, so obtained suitable task. Through experiments benchmark dataset, show proposed method improves over baseline both quantitatively qualitatively.