Reproduce “Closed-Form Factorization of Latent Semantics in GANs” 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

Yujun Shen, Bolei Zhou

SeFa, a closed-form factorization algorithm, efficiently identifies interpretable semantic dimensions within GAN latent spaces by directly analyzing pre-trained generator weights. This unsupervised method achieves comparable qualitative performance to supervised techniques for attribute editing and discovers a broader range of semantics in under one second per model, applicable across diverse GAN architectures and datasets.

Key insights

  • Interpretable semantic dimensions in GANs can be directly revealed by decomposing the internal architecture (pre-trained weights) of the generator.
  • The initial affine transformation within a GAN generator encodes critical information about how latent changes translate into significant image variations.
  • Eigenvalue decomposition of the weight matrix's transpose multiplied by itself (A^T A) provides a closed-form, computationally efficient solution to discover these principal semantic directions.

Summary by alphaXiv.

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