This repository implements GANSpace, a method for discovering interpretable latent directions in pre-trained Generative Adversarial Networks. The approach uses Principal Components Analysis on activation spaces to identify directions corresponding to interpretable attributes like viewpoint, aging, lighting, and time of day. It includes modified versions of BigGAN, StyleGAN, and StyleGAN2 that support per-layer latent vectors and provides interactive tools for exploring and applying these discovered controls to generated images.