A Comparison between AttnGAN and DF GAN: Text to Image Synthesis

Abstract

Nowadays conversion from text to high resolution image is a challenging task due to its wide variety of application area. For text to image conversion almost all systems use Generative Adversarial Networks as the basic part of the system and GAN guarantees semantic consistency between the text input and the generated image output. In this paper we are comparing two algorithms that is used for generating image from text. The first algorithm is the AttnGAN and the second one is the DF-GAN. AttnGAN builds on top of StackGAN by using attention network which allows it to capture word level information along with the broader sentence level information. The second algorithm is the DF-GAN, which uses single generator and discriminator model to synthesize high resolution images and also uses Matching-Aware Gradient Penalty (MA-GP) to get real images with real description. The model contains a Deep text-image Fusion Block (DFBlock) to generate image features from text. Both algorithms work efficiently for image generation from text but DF-GAN generates the perfect output than AttnGAN. The AttnGAN always focus on the textual part to generate output image but DF-GAN also focuses on background of image.