Celebrity Face Generation from Textual Descriptions

Abstract

This paper details the creation of a new text and image paired data set from the CelebA image and attribute collection in order to facilitate a text-to-face image generation model. It also details the design, development, and evaluation of a text-conditioned Generative Adversarial Deep Convolutional Network. The most relevant steps taken in the project consist in assembling meaningful and accurate text captions for the CelebA dataset, the use of GloVe and EMLo embeddings, and recreating the GAN-CLS architecture in TensorFlow given its original paper and Pytorch Repository. The goal of this project is to generate text conditioned faces as accurately and as ”real” as possible and to gain a better insight into the effects of specific text embeddings on the final synthetic images.