Research map / Image generation / Text-to-image diffusion
Finetuning Text-to-Image Diffusion Models for Fairness
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.
Frames text-to-image diffusion fairness as distributional alignment, combining a classifier-based distributional alignment loss with an adjusted direct finetuning of the sampling process to debias gender, race, and age in Stable Diffusion v1-5.
Problem
Text-to-image diffusion models generate skewed demographic distributions (e.g., gender, race) for occupational prompts, and existing debiasing methods are limited to prompting or approximate removal and cannot control the target distribution.
Approach
A distributional alignment loss (DAL) uses pretrained attribute classifiers and optimal-transport-derived dynamic targets to steer generated face attributes toward a user-defined distribution, plus an adjusted direct finetuning that sets the per-step gradient coefficients A_t=1 and B_t=I to avoid exploding and high-variance gradients through the sampling chain.
Key results
- Debiasing gender, race, and intersectional biases reduces gender bias from 0.67 to 0.16 and G.xR. bias from 0.21 to 0.06 on SD v1-5 test occupations, while keeping CLIP-T at 0.39.
- Finetuning only five soft tokens as a prompt prefix reduces gender bias from 0.67 to 0.24.
- Aligning age to 75% young / 25% old while debiasing gender and race yields 24.8% old individuals on average, with gender bias 0.15 and race bias 0.09.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Debias gender using distributional alignment and adjusted DFT | Stable Diffusion 1.5 · Occupational prompts | gender bias: .67\graysmall{$\pm$.29} -> .23{ $\pm$.16} | Helped |
| Debias race using distributional alignment and adjusted DFT | Stable Diffusion 1.5 · Occupational prompts | race bias: .42\graysmall{$\pm$.06} -> .12{ $\pm$.05} | Helped |
| Jointly debias gender and race | Stable Diffusion 1.5 · Occupational prompts | gender × race bias: .21\graysmall{$\pm$.03} -> .06{ $\pm$.02} | Helped |
| Finetune five soft tokens as a prompt prefix instead of the full text encoder (ablation) | Stable Diffusion 1.5 gender debiasing · Occupational prompts | gender bias: .23\graysmall{$\pm$.16} -> .24\graysmall{$\pm$.19} | No clear effect |
| Finetune U-Net instead of the text encoder (ablation) | Stable Diffusion 1.5 gender debiasing · Occupational prompts | CLIP-I: .77\graysmall{$\pm$.15} -> .90\graysmall{$\pm$.09} | Helped |
| Finetune U-Net instead of the text encoder (ablation) | Stable Diffusion 1.5 gender debiasing · Occupational prompts | gender bias: .23\graysmall{$\pm$.16} -> .22\graysmall{$\pm$.14} | No clear effect |
| Add U-Net finetuning to text encoder finetuning (ablation) | Stable Diffusion 1.5 gender debiasing · Occupational prompts | gender bias: .23\graysmall{$\pm$.16} -> .17\graysmall{$\pm$.13} | Helped |
| Add U-Net finetuning to text encoder finetuning (ablation) | Stable Diffusion 1.5 gender debiasing · Occupational prompts | CLIP-I: .77\graysmall{$\pm$.15} -> .80\graysmall{$\pm$.14} | Hurt |
| Add age alignment to joint gender and race debiasing (ablation) | Stable Diffusion 1.5 · Occupational prompts | frequency of age=old: .147\graysmall{$\pm$.216} -> .248\graysmall{$\pm$.091} | Helped |
| Add age alignment to joint gender and race debiasing (ablation) | Stable Diffusion 1.5 · Occupational prompts | gender bias: .16\graysmall{$\pm$.13} -> .15\graysmall{$\pm$.12} | No clear effect |
| Use adjusted DFT rather than naive DFT | Stable Diffusion diffusion sampling process · Fixed-target image semantics preservation optimization | training loss | Helped |
Surprising findings
- Naive direct finetuning of the diffusion sampling process fails to reduce the training loss, because the exact gradient through the sampling chain explodes and has high variance at large time steps.
- Finetuning the U-Net can deteriorate facial skin texture and make the model overfit to the classifier, producing images whose predicted gender disagrees with human perception.
Limitations
- Only binary gender and four consolidated race categories are considered; non-binary and mixed identities are not addressed.
- Classifiers trained on faces define groups, which may neglect non-facial characteristics; debiasing generalizes only partially to general prompts (e.g., 'woman' prompts show limited Black and Indian representation gains).
- Finetuning reduces naturalness and smoothness of some generated images, as judged by the lead author's subjective assessment.
Experimental setup
Datasets: CelebA FairFace LAION-Aesthetics V2 DiffusionDB
Benchmarks: occupational prompts test set (50 occupations) non-templated LAION-Aesthetics prompts (40) multi-face prompts sports, styled occupation, and personal descriptor prompts
Metrics: gender bias racial bias gender x race bias CLIP-T CLIP-I DINO similarity age=old frequency
Baselines: debiasVL Ethical Intervention Concept Algebra UCE Original SD
Builds on: Stable Diffusion DPM-Solver++ LoRA DINOv2 textual inversion
Closest papers by meaning
- FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.92)
- Leveraging Diffusion Perturbations for Measuring Fairness in Computer Vision (AAAI 2024 · similarity 0.91)
- Fair Text-to-Image Diffusion via Fair Mapping (AAAI 2025 · similarity 0.91)
- AIM-Fair: Advancing Algorithmic Fairness via Selectively Fine-Tuning Biased Models with Contextual Synthetic Data (CVPR 2025 · similarity 0.88)
- SocialCounterfactuals: Probing and Mitigating Intersectional Social Biases in Vision-Language Models with Counterfactual Examples (CVPR 2024 · similarity 0.88)
- Implicit Bias Injection Attacks against Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.88)
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.