RAT-Cycle-GAN for text to image

Abstract

We present a method that enhances the RATGAN using the idea of CycleGAN to prevent the model from collapsing. The generator with the proposed recurrent affine transformation for text-to-image synthesis and the fusion blocks are connected by an RNN during the generation of fake images to assure the global assignment of text information. The transformation from the phony images to the text is then accomplished using a CNN and encoder with LSTM. A comparison of the generated text with the original text ensures the generator’s preciseness. In order to achieve the state of the art, we evaluated both unconditional and conditional generation’s effectiveness in relation to the accepted standards.