Research map / Image generation / GANs
FairStyle: Debiasing StyleGAN2 with Style Channel Manipulations
Cemre Efe Karakas, Alara Dirik, Eylül Yalçınkaya, Pinar Yanardag
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 changed | Where | Measured | Effect |
|---|---|---|---|
| debias StyleGAN2 with FairStyle style-channel bias tensors (vs pretrained) | StyleGAN2 · FFHQ Gender+Glasses | KL divergence to uniform: 0.2495 -> 2.41\times10^{-4} | Improved |
| debias StyleGAN2 for single attribute Glasses | StyleGAN2 · FFHQ Glasses | KL divergence to uniform: 0.180 -> 0 | Improved |
| debias Black attribute with CLIP-based FairStyle | StyleGAN2 · FFHQ Black (race) | KL divergence to uniform: 0.603 -> 8.00 \times 10^{-6} | Improved |
| debias Asian attribute with CLIP-based FairStyle | StyleGAN2 · FFHQ Asian (race) | KL divergence to uniform: 0.319 -> 7.20 \times 10^{-7} | Improved |
| use FairStyle instead of FairGen for Age+Gender | FairStyle vs FairGen · FFHQ Age+Gender | KL divergence to uniform: 3.73\times10^{-2} -> 2.57\times10^{-2} | Improved |
| debias Smiling with FairStyle instead of StyleFlow | FairStyle vs StyleFlow · FFHQ Smiling | KL divergence to uniform: 0.045 -> 8\times10^{-8} | Improved |
| StyleFlow vs pretrained StyleGAN2 on Smiling | StyleFlow · FFHQ Smiling | KL divergence to uniform: 0.011 -> 0.045 | Hurt |
| debias with FairStyle (image quality cost) | StyleGAN2 · FFHQ debiased sets | FID: 14.11 -> 14.72 | Hurt |
Later papers that compared with this method
1 later accepted paper reports FairStyle as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Generative Visual Prompt: Unifying Distributional Control of Pre-Trained Generative Models (NeurIPS 2022 · better on 2 of 3 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- StyleGAN2 Distillation for Feed-forward Image Manipulation (ECCV 2020 · similarity 0.88)
- Analyzing and Improving the Image Quality of StyleGAN (CVPR 2020 · similarity 0.87)
- StyleSpace Analysis: Disentangled Controls for StyleGAN Image Generation (CVPR 2021 · similarity 0.87)
- A Geometrical Perspective on Image Style Transfer with Adversarial Learning (TPAMI 2020 · similarity 0.86)
- Hierarchical Semantic Regularization of Latent Spaces in StyleGANs (ECCV 2022 · similarity 0.86)
- MOST-GAN: 3D Morphable StyleGAN for Disentangled Face Image Manipulation (AAAI 2022 · similarity 0.86)
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.
