Semantic layout aware generative adversarial network for text-to-image generation

Text-to-image(T2I) generation methods aim to synthesize a high-quality image which is semantically consistent with the given text descriptions. Previous (T2I) generative adversarial networks generally first create a low-resolution image with rough shapes and colors, and then refine the initial image into a high-resolution image. Most stacked architecture still remains two main problems. (1) The final images generated by these methods depend heavily on the quality of the initial image. If the initial one is not initialized correctly, the resulted image seems like a simple combination of visual features from several images scales. (2) The cross-modal fusion methods about text and image that previous works widely adopted is limited in the text-image fusion process. In the paper, we propose a novel generation model, which introduce a one-stage backbone directly generate high-quality images without multi generators and a novel semantic layout deep fusion network to sufficiently fuse text features and image features. Experiments on the challenging CUB and COCO-Stuff datasets demonstrates the ability of our model in generating images, regarding both semantic consistency with input text description and visual fidelity.