FaceForge: Sculpting Realistic Faces from Detailed Textual Prompts

Within the domain of generative adversarial networks (GANs), notable progress has been made in generating lifelike visuals based on textual stimuli. Nevertheless, the primary emphasis has centered on basic assignments such as producing visual representations of flowers using textual descriptions. This work aims to explore the unexplored field of face generation, specifically focusing on generating faces from detailed textual descriptions such as "A person with wavy hair, an oval face, and a moustache." Motivated by the potential influence on important applications such as criminal face reconstruction, we address the problem of limited datasets by developing an algorithm that automatically converts. We utilise the advanced DC-GAN with GAN-CLS loss to address the complexities arising from subtle information and varied caption lengths. In order to enhance the resilience of the training process, we propose the incorporation of label inversion for both real and synthetic images, along with the introduction of noise into the discriminator.