Extracting Semantic Knowledge from GANs with Unsupervised Learning
Jianjin Xu1,2 Zhaoxiang Zhang3 Xiaolin Hu2
1Panzhihua University
2Tsinghua University
3Chinese Academy of Sciences

[Online Demo] [Paper] [Code]

We propose K-means with Linear Separability Heuristic (KLiSH) to decode semantic parts in GAN's features. No human annotated semantic segmentation is needed throughout the process. The decoded semantic parts are fine-grained and accurate. One can use the decoded semantics for downstream semantic segmentation and semantic image editing tasks.

Work done while working at THBI, Tsinghua as a Research Assistant.