AI Paper Atlas

Research map / Image generation / GANs

Extracting Semantic Knowledge From GANs With Unsupervised Learning

Jianjin Xu, Zhaoxiang Zhang, Xiaolin Hu

TPAMI 2023 · Image generation › GANs · approach: Large-scale GAN latent analysis · 22 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use KLiSH clustering instead of AHC (arccos)GAN feature clustering · StyleGAN2-FFHQ, K=26MIoU: 43.8 -> 56.3Improved
use KLiSH instead of K-meansGAN feature clustering · StyleGAN-CelebAHQ, K=30MIoU: 38.5 -> 45.7Improved
use KLiSH instead of AHC (arccos)GAN feature clustering · PGGAN-CelebAHQ, K=26AMI: 36.0 -> 51.0Improved
use KLiSH instead of AHC (arccos)GAN feature clustering · StyleGAN2-FFHQ, K=26ARI: 69.2 vs 65.4Hurt
use KLiSH labels instead of AHC labels for unsupervised segmentationUFGS · CelebAMask-HQmIoU: 36.6 -> 47.2Improved
compare zero-annotation UFGS against one-shot LSEUFGS (KLiSH) · CelebAMask-HQmIoU: 47.2 vs 47.1No clear effect
compare UFGS against 10-shot LSEUFGS (KLiSH) · CelebAMask-HQmIoU: 47.2 vs 58.5Hurt
replace SVM with softmax-trained LSE (classical linear separability) (ablation)feature segmentation probe · StyleGAN2-FFHQmIoU: 81.7 vs 82.7No clear effect
replace SVM with LSE (ablation)feature segmentation probe · StyleGAN-BedroommIoU: 30.9 vs 37.3Hurt
use centroids of features instead of linear classifier (ablation)K-means-style clustering assumption · StyleGAN2-FFHQmIoU: 82.7 -> 49.6Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy