Imaginface : Text-Driven Human Face Generation

Abstract

In the realm of text-to-image synthesis, generating realistic human faces from textual descriptions using Generative Adversarial Networks (GANs) poses significant challenges such as complex designs, reliance on additional networks for semantic alignment, and high computational costs. Our approach employed a Deep Convolutional GAN (DC-GAN) framework, utilizing the CelebA dataset to enhance training, and incorporated a CNN-RNN model to align textual descriptions with facial features, achieving a Fr�echet Inception Distance (FID) score of 15.32. We also investigated the Deep Fusion GAN (DF-GAN), a one-stage architecture that simplifies high-resolution image synthesis and improves text-image semantic consistency with its Target-Aware Discriminator, Matching Aware Gradient Penalty, and One-Way Output. DF-GAN demonstrated superior performance with an FID score of 12.05 on the CUB 200 2011 dataset but recorded a higher FID score of 45.03 on the CelebA dataset, highlighting dataset-specific challenges.