Research map / Image generation / Text-to-image diffusion
On Measuring Fairness in Generative Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| replace Baseline mean SA-classifier output with CLEAM point estimate | fairness measurement of StyleGAN2 · GenData-StyleGAN2 Gender, ResNet-18 | point error e_mu: 4.98 -> 0.62 | Helped |
| use CLEAM instead of Baseline | fairness measurement, average over classifiers · StyleGAN2 and StyleSwin GenData | average point error: 8.23% -> 1.24% | Helped |
| use CLEAM interval estimate instead of Baseline confidence interval | fairness measurement · StyleGAN2 Gender | average interval max error e_rho: 9.70% -> 1.99% | Helped |
| use CLEAM instead of Baseline with CLIP classifier | fairness measurement of Stable Diffusion · GenData-SDM Gender prompts | average point error: 5.75% -> 0.44% | Helped |
| use CLEAM interval estimate for SDM | fairness measurement of Stable Diffusion · GenData-SDM Gender prompts | average interval error: 8.74% -> 3.61% | Helped |
| use Diversity measure (VGG-16) instead of CLEAM | fairness measurement · StyleGAN2 and StyleSwin GenData | point error: 14.30% -> 0.90% | Hurt |
| use a less accurate classifier (VGG-16 vs ResNet-18) for Baseline (ablation) | Baseline fairness measurement · StyleGAN2 Gender | point error: 4.98 -> 17.13 | Hurt |
| use a less accurate classifier for CLEAM (ablation) | CLEAM · StyleGAN2 Gender | point error: 0.62 -> 0.93 | No clear effect |
| add Black Box Shift Estimation (BBSE) correction (ablation) | Baseline measurement · GenData classifier setup | point error: 4.20% <= e <= 3.38% | Mixed |
| add Black Box Shift Correction (BBSC) (ablation) | Baseline measurement · GenData classifier setup | point error | No clear effect |
| vary pseudo-generator bias p0 from 0.5 to 0.9 with CLEAM (ablation) | CLEAM · CelebA Gender/BlackHair, AFHQ Cat/Dog | average point error: 6.93% -> 0.75% | Helped |
| use CLEAM at unbiased point p0=0.5 with uniformly inaccurate classifier (ablation) | CLEAM · pseudo-generator Gender, BlackHair ResNet-18 | point error | Hurt |
Closest papers by meaning
- Distributionally Generative Augmentation for Fair Facial Attribute Classification (CVPR 2024 · similarity 0.87)
- AIM-Fair: Advancing Algorithmic Fairness via Selectively Fine-Tuning Biased Models with Contextual Synthetic Data (CVPR 2025 · similarity 0.86)
- FairDD: Fair Dataset Distillation (NeurIPS 2025 · similarity 0.85)
- FairNet: Dynamic Fairness Correction without Performance Loss via Contrastive Conditional LoRA (NeurIPS 2025 · similarity 0.85)
- Intersectional Unfairness Discovery (ICML 2024 · similarity 0.84)
- Fairness-Aware Estimation of Graphical Models (NeurIPS 2024 · similarity 0.84)
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.