Research map / Image generation / Fast diffusion sampling
Debiasing Pretrained Generative Models by Uniformly Sampling Semantic Attributes
Walter Gerych, Kevin Hickey, Luke Buquicchio, Kavin Chandrasekaran, Abdulaziz Alajaji, Elke Rundensteiner, Emmanuel Agu
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 |
|---|---|---|---|
| Apply Distribution Mapping to the generator | Progressive GAN · CelebA-HQ Race | KL divergence from uniform: 0.68 -> 0.41 | Improved |
| Apply Distribution Mapping to the latent diffusion model | LDM · CelebA-HQ Age | KL divergence from uniform: 0.411 -> 0.130 | Improved |
| Use Latent Editing instead of vanilla sampling | LDM · CelebA-HQ Age | KL divergence from uniform: 0.411 -> 0.525 | Hurt |
| Use Latent Editing instead of vanilla sampling | Progressive GAN · CelebA-HQ Race | KL divergence from uniform: 0.68 -> 0.62 | Improved |
| Use Polarity antimode sampling (rho=2) instead of standard sampling | Progressive GAN · CelebA-HQ Race | KL divergence from uniform: 0.68 -> 0.49 | Improved |
| Use Polarity mode sampling (rho=-2) instead of standard sampling | Progressive GAN · CelebA-HQ Race | KL divergence from uniform: 0.68 -> 0.57 | Improved |
| Use MaGNET instead of standard sampling | Progressive GAN · CelebA-HQ Race | KL divergence from uniform: 0.68 -> 0.56 | Improved |
| Use Shifted Semantic Distribution instead of standard sampling | Progressive GAN · CelebA-HQ Race | KL divergence from uniform: 0.68 -> 0.60 | Improved |
| Add bias correction (Algorithm 2) when building the Fair Noise dataset with a biased linear classifier (ablation) | Fair Noise dataset construction · Shapes (circle/square) | group density (ideal 0.5) | Improved |
| Add bias correction (Algorithm 2) with a weakened biased classifier (ablation) | Fair Noise dataset construction · UTKFace DCGAN, 10 age bins | group density (ideal 0.1) | Improved |
| Use Latent Editing with biased classifier rather than perfect classifier (ablation) | Latent Editing · Shapes | group density (ideal 0.5) | Hurt |
| Use Distribution Mapping vs baselines on age debiasing | Distribution Mapping · UTKFace DCGAN Age | age distribution uniformity | Mixed |
Closest papers by meaning
- Balancing Act: Distribution-Guided Debiasing in Diffusion Models (CVPR 2024 · similarity 0.90)
- Diverse Yet Biased: Towards Mitigating Biases in Generative AI (Student Abstract) (AAAI 2024 · similarity 0.89)
- FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.88)
- Would Deep Generative Models Amplify Bias in Future Models? (CVPR 2024 · similarity 0.88)
- Low-Biased General Annotated Dataset Generation (CVPR 2025 · similarity 0.87)
- Debiasing Synthetic Data Generated by Deep Generative Models (NeurIPS 2024 · similarity 0.87)
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.
