Third Time’s the Charm? Image and Video Editing with StyleGAN3

AbstractStyleGAN is arguably one of the most intriguing and well-studied generative models, demonstrating impressive performance in image generation, inversion, manipulation. In this work, we explore recent StyleGAN3 architecture, compare it to its predecessor, investigate unique advantages, as well drawbacks. particular, demonstrate that while can be trained on unaligned data, still use aligned data for training, without hindering ability generate imagery. Next, our analysis disentanglement different latent spaces indicates commonly used \(\mathcal {W}/\mathcal {W}+\) are more entangled than their StyleGAN2 counterparts, underscoring benefits using StyleSpace fine-grained editing. Considering observe existing encoder-based techniques struggle when data. We therefore propose an encoding scheme solely yet invert images. Finally, introduce a novel video inversion editing workflow leverages capabilities fine-tuned generator reduce texture sticking expand field view edited video. KeywordsGenerative Adversarial NetworksImage