Rewriting geometric rules of a GAN

Deep generative models make visual content creation more accessible to novice users by automating the synthesis of diverse, realistic based on a collected dataset. However, current machine learning approaches miss key element creative process - ability synthesize things that go far beyond data distribution and everyday experience. To begin address this issue, we enable user "warp" given model editing just handful original outputs with desired geometric changes. Our method applies low-rank update single layer reconstruct edited examples. Furthermore, combat overfitting, propose latent space augmentation style-mixing. allows create synthesizes endless objects defined changes, enabling new without burden curating large-scale We also demonstrate can be composed achieve aggregated effects, present an interactive interface through composition. Empirical measurements multiple test cases suggest advantage our against recent GAN fine-tuning methods. Finally, showcase several applications using models, including interpolation image editing.