Facial Image Generation using fully trained Generative Adversarial Network

Recent advancements in facial image generation have been significant. However, many existing methods are constrained to generating face images solely from random noise, lacking the capability to synthesize images based on specific features. In this work, the issue of face synthesis from features is mapped, trying to produce facial images that have unique features that match pre-specified criteria. In order to accomplish this goal, this research presents new techniques for generating attribute-aware face images using Generative Adversarial Networks (GANs), which include Conditional GAN (CGAN), Wasserstein GAN (WGAN), and Deep Convolutional GAN (DCGAN).This paper’s method starts by converting the input text into an array of its features and then mapped with GAN. Subsequently, three stacked generators are utilized to produce face images of resolutions 64×64, 128×128, and 256×256, respectively, using characteristic traits as the input. Furthermore, a proposed image-attribute matching loss is made to improve the relation between the input vector with the generated images. Prolonged testing on the CelebA dataset confirms the efficacy of the CGAN, WGAN, and DCGAN methods presented in this work. This paper shows the comparison of CGAN, DCGAN and WGAN based on various performance evaluation parameters such as Inception score, Generator’s loss and Discriminators loss for a particular number of training epochs.

I. Introduction

In recent years, Interest in computer vision has noticeably increased, especially in the domain of automatically creating face images with realistic features through computational techniques.