Comparative Analysis of AttnGAN, DF-GAN and SSA-GAN

In computer vision, generating visuals based on the associated text is important, demanding, and interesting task. GANs (Generative Adversarial Networks) are the most powerful generative models for computer vision and natural language processing. GANs ensure that the synthesized image matches the input text semantically. The purpose of this paper is to compare three algorithms for producing images from text. These include Attentional Generative Adversarial Networks (AttnGAN), Deep-Fusion Generative Adversarial Networks (DF-GAN) and Semantic-Spatial Aware Generative Adversarial Networks (SSA-GAN). AttnGAN extends StackGAN++ by including attention in a multi-stage refining pipeline. DF-GAN, and SSA-GAN on the other hand are based on one stage architecture. DF-GAN incorporates Deep Fusion Block (DFBlock) that effectively fuses textual and visual information whereas SSA-GAN uses novel Semantic-Spatial Aware Convolutional Network (SSACN) block for efficient and deep text and image features fusion.