Dialog Driven Face Construction using GANs

This paper presents an end-to-end pipeline using Generative Adversarial Networks (GANs) for face construction based on speech-based descriptions, and iterative editing of the generated image to arrive at a close approximation of the expected face. We propose a dialog-based interaction with the system where the user and system take turns providing descriptions and generating images respectively. A rule-based Natural Language (NL) Parser is used to extract facial attribute descriptors from textual descriptions, MSG-Style GAN (Multi-Scale Gradient Style GAN) for face generation, and Attribute GAN (AttGAN) for facial attribute manipulation.