SeFa - Closed-Form Factorization of Latent Semantics in GANs

Figure: Versatile semantics found from various types of GAN models using SeFa.

Closed-Form Factorization of Latent Semantics in GANs
Yujun Shen, Bolei Zhou
arXiv preprint arXiv:2007.06600

[Paper] [Project Page] [Demo]

In this repository, we propose a closed-form approach, termed as SeFa, for unsupervised latent semantic factorization in GANs. With this algorithm, we are able to discover versatile semantics from different GAN models trained on various datasets. Most importantly, the proposed method does not rely on pre-trained semantic predictors and has an extremely fast implementation (i.e., less than 1 second to interpret a model). Below show some interesting results on anime faces, cats, and cars.

NOTE: The following semantics are identified in a completely unsupervised manner, and post-annotated for reference.

Anime Faces
Pose	Mouth	Painting Style

Cats
Posture (Left & Right)	Posture (Up & Down)	Zoom

Cars
Orientation	Vertical Position	Shape

BibTeX

@article{shen2020closedform,
  title   = {Closed-Form Factorization of Latent Semantics in GANs},
  author  = {Shen, Yujun and Zhou, Bolei},
  journal = {arXiv preprint arXiv:2007.06600},
  year    = {2020}
}

Code Coming Soon