Reproduce “WarpedGANSpace: Finding non-linear RBF paths in GAN latent space” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Christos Tzelepis, Georgios Tzimiropoulos, Ioannis Patras

WarpedGANSpace presents an unsupervised method for discovering non-linear paths in GAN latent spaces using Radial Basis Function (RBF) warping. This approach enables more natural and pronounced attribute transformations, achieving up to 99.8% discriminability for discovered transformations and yielding 43.66 degrees of yaw rotation on StyleGAN2, compared to 18.93 degrees for linear methods.

Key insights

  • Non-linear paths, derived from RBF gradients, enable more natural and context-aware semantic transformations in GAN latent spaces.
  • An adversarial training dynamic between warping and reconstructor networks can discover distinct and interpretable non-linear directions without supervision.
  • Objective quantitative evaluation using pre-trained attribute classifiers is crucial for rigorously comparing and validating GAN interpretability methods, moving beyond subjective visual assessment.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.