A GAN model to produce Photo realistic Images via text command.

Abstract

In today’s era of Computer vision, Imagination can come true. There is a wide range of commercial applications for generating images via text commands. It’s just like you tell your computer to draw an art and your computer does it. This challenging problem has a solution with two staged Generative Adversarial Network (GAN) model. In this paper we propose two staged Generative Adversarial Network to generate a photo realistic image. The text command is the input for the first stage GAN, which outputs a very basic image with almost no resolution. This almost no resolution image undergoes sketch refinement. This output of first stage GAN and text description is fed as an input to the second stage of GAN. Second stage GAN generates high resolution photo realistic image. The second stage GAN rectifies the defective output image of first stage GAN recursively. We use the augmentation technique to ensure the smoothness of the image. Wide ranging experiments shows two staged GAN has a prominent growth on creating photo realistic images based on text description.