Text-to-image generation based on AttnDM-GAN and DMAttn-GAN: applications and challenges

Abstract

The deep fake faces generation using generative adversarial networks (GANs) has reached an incredible level of realism where people can’t differentiate the real from the fake. Text-to-face is a very challenging task compared to other text-to-image syntheses because of the detailed, precise, and complex nature of the human faces in addition to the textual description details. Providing an accurate realistic text-to-image model can be useful for many applications such as criminal identification where the model will be acting as the forensic artist. This paper presents text-to-image generation based on attention dynamic memory (AttnDM-GAN) and dynamic memory attention (DMAttn-GAN) that are applied to different datasets with an analysis that shows the different complexity of different datasets’ categories, the quality of the datasets, and their effect on the results of the resolution and consistency of the generated images.