Research map / Image generation / Text-to-image diffusion
DRPO: Disentangling Demographic Bias from Rewards for Fair Diffusion Alignment
YeonGyu Han, Junah Jung, Dongheon Lee
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 |
|---|---|---|---|
| add demographic-direction projection to GRPO reward | DanceGRPO · SDXL neutral prompts | Race entropy: 0.89 -> 1.67 | Improved |
| add demographic-direction projection to GRPO reward | DanceGRPO · SDXL neutral prompts | HPS: 27.42 -> 27.31 | No clear effect |
| add projection to Diffusion-DPO reward | Diffusion-DPO · SDXL neutral prompts | Race entropy: 1.08 -> 1.61 | Improved |
| add diversity incentives (DRIFT) to GRPO | DanceGRPO · SDXL neutral prompts | Race entropy: 0.89 -> 0.93 | No clear effect |
| add diversity incentives (DRIFT) to GRPO | DanceGRPO · SDXL neutral prompts | Vendi Score: 0.34 -> 0.73 | Improved |
| add ITI-GEN inclusive prompts to GRPO | DanceGRPO · SDXL neutral prompts | HPS: 27.42 -> 26.84 | Hurt |
| increase projection dimension k from 1 to 5 (ablation) | DRPO (w/ GRPO) · SDXL | Race entropy: 1.34 -> 1.67 | Improved |
| increase projection dimension k from 5 to 10 (ablation) | DRPO (w/ GRPO) · SDXL | HPS: 27.31 -> 27.14 | Hurt |
| use feature-level projection instead of output-level group-mean subtraction (ablation) | DRPO (w/ GRPO) · SDXL | Race entropy: 1.23 -> 1.67 | Improved |
| use CLIP zero-shot instead of FairFace labels for probe (ablation) | DRPO (w/ GRPO) · SDXL | Race entropy: 1.67 -> 1.59 | No clear effect |
| degrade demographic classifier accuracy to 80% (ablation) | DRPO (w/ GRPO) · SDXL | Race entropy: 1.67 -> 1.56 | No clear effect |
| compare DRPO with DanceGRPO in human 2AFC | DRPO (w/ GRPO) · Human evaluation | preference rate: 47.3% | No clear effect |
Closest papers by meaning
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 2024 · similarity 0.90)
- Calibrated Multi-Preference Optimization for Aligning Diffusion Models (CVPR 2025 · similarity 0.90)
- Smoothed Preference Optimization via ReNoise Inversion for Aligning Diffusion Models with Varied Human Preferences (ICML 2025 · similarity 0.90)
- Aligning Diffusion Models by Optimizing Human Utility (NeurIPS 2024 · similarity 0.89)
- A Dense Reward View on Aligning Text-to-Image Diffusion with Preference (ICML 2024 · similarity 0.89)
- Diffusion-NPO: Negative Preference Optimization for Better Preference Aligned Generation of Diffusion Models (ICLR 2025 · similarity 0.89)
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.
