Web Server Based StackGAN-v2 Implementation

Abstract

Implementation of StackGAN on a website with the help of a web server to generate unique pictures based on a description given by a user. Computer vision generating good quality images from the user's text description is a troublesome problem and has many useful applications. Samples from current Text to Image methods can essentially demonstrate comprehension of the explanations given, but they fail to provide the information needed and the vivid sections of the objects. With this paper, we use Stacked Generative Adversarial Networks (StackGAN) to get realistic image using the text descriptions given by the user. Via a method of sketch-refinement we break the hard issue down into more manageable subproblems. We have used StackGAN because StackGAN-v2 shows more robust training behavior than StackGAN-v1 by approximating several distributions together. In StackGAN images are generated from different branches of the tree at multiple scales corresponding to the same scene. Comprehensive tests and comparisons with benchmark data sets indicate that significant changes are made by the suggested approach in generating photorealistic images based on text descriptions.

 Keywords: Text to Image; StackGAN-v2; GANs