Discovering and Modifying Latent Directions of Synthetic Images while Maintaining Image Realism

Document Type


Degree Name

Master of Science (MS)


Computer Science and Info Sys

Date of Award

Fall 2021


In recent years, Generative Adversarial Networks (GANs) have made immense progress in their ability to generate high quality synthetic images. However, even though these generators can now produce high resolution images, modifying synthetic images whilst also preserving their realism has not developed as rapidly. In this paper, we will be studying the issue of poor synthetic image modifications. We will begin this by looking at the latent space that is sampled to generate the images as well as attempting to identify and modify the latent space of the image generated by a GAN (which changes certain important features of the images). By doing so we can create a set of controls that will let us directly modify specific aspects of the images while keeping the resolution optimal. We will also create an architecture that can generate synthetic images that look like a given input image.


Omar El Ariss

Subject Categories

Computer Sciences | Physical Sciences and Mathematics