*In this project, we apply the StyleSpace analysis on StyleGAN, StyleALAE and StyleFlow architectures to study the effect of the mapping network on the disentanglement properties observed in the latent spaces W and W+, and the Style Space S. Furthermore, we study how these spaces compare to a disentangled space created using an invertible interpretation network from [4]. We compare the image generation quality and disentanglement between these models as well as the manipulation quality between Style Space S and StyleFlow. *
This code reproduces the results in the accompanying project report. Please follow the steps listed in the README files in the respective folders to reproduce the results for the respective models.