Research map / Image generation / Text-to-image diffusion
FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FairImagen, a training-free post-hoc debiasing method that projects CLIP/SD prompt embeddings with Fair PCA, adds empirical noise, and uses a Cartesian-product joint projection to reduce gender and race bias in text-to-image generation.
Problem
Text-to-image diffusion models replicate and amplify demographic biases (gender, race, intersectional), and existing post-hoc prompt-embedding debiasing methods are limited to binary single-attribute directions and tend to cause semantic drift and loss of prompt fidelity.
Approach
Apply Fair PCA to pooled and token-level prompt embeddings to remove group-dependent directions while keeping semantic variance, add empirically sampled noise along group bias directions to avoid over-neutral outputs, and debias multiple attributes jointly via a Cartesian-product group space.
Key results
- Gender debiasing: FairImagen reaches fairness 0.56 vs 0.167 for Base and 0.507 for SDID, with Avg 0.624 (best among post-hoc methods), though accuracy (0.771) and MUSIQ (0.541) lag some baselines.
- Joint gender+race debiasing: FairImagen gets Avg 0.611 versus 0.508 for Base and 0.574 for TBIE, with gender fairness 0.537 and race fairness 0.32.
- FairPrompt (human-designed per-image prompts) achieves the highest fairness across settings (e.g., 0.732 gender fairness), serving as an upper bound rather than a fair post-hoc comparison.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the sampled empirical noise with mean empirical magnitude (ablation) | FairImagen · Gender debiasing | Gender Fairness: 0.455 -> 0.432 | Hurt |
| replace empirical noise with fixed directional noise (ablation) | FairImagen · Gender debiasing | Gender Fairness: 0.455 -> 0.258 | Hurt |
| replace empirical noise with random Gaussian noise (ablation) | FairImagen · Gender debiasing | Gender Fairness: 0.455 -> 0.167 | Hurt |
| replace empirical noise with a fixed random Gaussian vector (ablation) | FairImagen · Gender debiasing | Gender Fairness: 0.455 -> 0.258 | Hurt |
| replace empirical noise with a constant bias shift (ablation) | FairImagen · Gender debiasing | Gender Fairness: 0.455 -> 0.242 | Hurt |
| remove FairPCA projection (ablation) | FairImagen · Gender debiasing | Gender Fairness: 0.455 -> 0.136 | Hurt |
| replace the joint demographic-space projection with stacked group indicators (ablation) | FairImagen · Gender + race debiasing | Gender Fairness: 0.429 -> 0.227 | Hurt |
| replace the joint demographic-space projection with sequential projections (ablation) | FairImagen · Gender + race debiasing | Gender Fairness: 0.429 -> 0.106 | Hurt |
| use the joint Cartesian-product demographic space instead of stacked or sequential projections (ablation) | FairImagen · Gender + race debiasing | Race Fairness: 0.345 -> 0.325 | Mixed |
Surprising findings
- FairPrompt introduces female representations even for historically male roles such as a middle ages blacksmith, whereas FairImagen preserves the male-leaning intent.
- Bypassing the projection (noise only off) gives high accuracy and MUSIQ but minimal fairness gain, showing the projection drives most of the debiasing.
Limitations
- Limited to binary gender and coarse race categories; non-binary and multi-ethnic identities not covered.
- Inherits biases of the CLIP encoder and cannot fully disentangle bias entangled with semantics.
- Fairness-utility trade-off requires empirical tuning of projection dimension and noise scale; evaluation limited to occupational prompts.
Experimental setup
Datasets: WinoBias Occupations list from Wikipedia
Benchmarks: WinoBias-extended occupations (120 professions)
Metrics: Fairness (normalized deviation score via DeepFace) Accuracy (CLIPScore x2.5) MUSIQ Avg
Baselines: FairPrompt ForcePrompt SAL CDA TBIE SDID SDID-AVG ITI-GEN FairQueue
Builds on: FairPCA SDID TBIE Stable Diffusion
Closest papers by meaning
- Fair Text-to-Image Diffusion via Fair Mapping (AAAI 2025 · similarity 0.93)
- Finetuning Text-to-Image Diffusion Models for Fairness (ICLR 2024 · similarity 0.92)
- Rethinking Training for De-biasing Text-to-Image Generation: Unlocking the Potential of Stable Diffusion (CVPR 2025 · similarity 0.90)
- FairRAG: Fair Human Generation via Fair Retrieval Augmentation (CVPR 2024 · similarity 0.90)
- LightFair: Towards an Efficient Alternative for Fair T2I Diffusion via Debiasing Pre-trained Text Encoders (NeurIPS 2025 · similarity 0.89)
- FairQueue: Rethinking Prompt Learning for Fair Text-to-Image Generation (NeurIPS 2024 · 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.