Research map / Image generation / Fast diffusion sampling
Leveraging Diffusion Perturbations for Measuring Fairness in Computer Vision
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.
Uses diffusion-based image generation and inpainting to build a race-balanced occupation dataset and a perturbation-based fairness metric, then benchmarks CLIP, FLAVA, and LENS models on occupation classification.
Problem
Computer vision and vision-language models lack demographically balanced evaluation datasets where images share backgrounds and context, making downstream fairness hard to measure.
Approach
Generate base occupation images with text-to-image diffusion, then inpaint only the person with different perceived race identifiers to create perturbation sets sharing identical context, filtered by VQA and a race attribution model, and measure fairness as the median per-set standard deviation of true-label probability.
Key results
- On the difficult label set, FLAVA achieves the highest fairness metric (0.983) and accuracy (90.1%), while CLIP-LAION400M has the lowest fairness (0.835).
- Human MTurk review of 4000 images gives 85.1% realism score and 91.0% race fidelity score overall.
- Non-Caucasian perceived groups show lower classification accuracy than Caucasian (Black: -6.09%, East Asian: -3.21%, Indian: -3.20%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use difficult labels instead of base labels | FLAVA · Occupation classification | classification accuracy: 92.0% -> 90.1% | No clear effect |
| use difficult labels instead of base labels | CLIP-LAION400M · Occupation classification | classification accuracy: 98.3% -> 75.1% | Hurt |
| use difficult labels instead of base labels | CLIP-LAION2B · Occupation classification | classification accuracy: 98.9% -> 79.8% | Hurt |
| use difficult labels instead of base labels | CLIP-OpenAI · Occupation classification | classification accuracy: 95.2% -> 68.6% | Hurt |
| use ViT-L/14 instead of ViT-B/32 (ablation) | CLIP models on difficult-label occupation classification · Occupation classification, difficult labels | fairness metric: 0.835 -> 0.862; 0.849 -> 0.908; 0.884 -> 0.903 | Mixed |
| specify balanced male and female gender in generation and inpainting prompts (ablation) | fairness evaluation on the four vision-language models · Occupation classification, gender-balanced sample | fairness metric ordering: FLAVA (0.944) > CLIP-OpenAI (0.874) > CLIP-LAION2B (0.827) > CLIP-LAION400M (0.821) | No clear effect |
| restrict to base images correctly classified by all evaluated models (ablation) | fairness evaluation on the four vision-language models · Occupation classification, difficult labels | fairness metric ordering: FLAVA (0.98), CLIP-OpenAI (0.884), CLIP-LAION2B (0.861), CLIP-LAION400M (0.835) | No clear effect |
| remove all negative labels except the top misclassified label (ablation) | fairness evaluation on the four vision-language models · Occupation classification, difficult labels | fairness metric ordering: FLAVA (0.999), CLIP-OpenAI (0.876), CLIP-LAION2B (0.866), CLIP-LAION400M (0.847) | No clear effect |
| use CLIP-OpenAI rather than CLIP-LAION400M | CLIP models · Occupation classification, difficult labels | fairness metric: 0.835 -> 0.884 | Helped |
| use CLIP-LAION2B rather than CLIP-LAION400M | CLIP models · Occupation classification, difficult labels | fairness metric: 0.835 -> 0.849 | Helped |
| use a non-Caucasian perceived identity label rather than Caucasian | evaluated vision-language models · Occupation classification, all occupations | classification accuracy: Black: -6.09%, East Asian: -3.21%, Indian: -3.20% | Hurt |
| use LENS H14-FlanT5-xl rather than LENS H14-FlanT5-xxl | LENS · Occupation classification, difficult labels | fairness metric: 0.837 -> 0.965 | Helped |
Surprising findings
- FLAVA, trained on much less data than CLIP models, achieves the highest accuracy and fairness on difficult labels, hypothesized to stem from a smaller vocabulary.
- CLIP-LAION2B has lower fairness than CLIP-OpenAI despite roughly 5x more training data, suggesting diminishing fairness returns with dataset size.
Limitations
- Generated images are not guaranteed to be balanced along other demographic traits such as gender; the dataset is skewed male for several occupations.
- Perceived race is a socially constructed label and the approach may reify a fixed definition of race.
- Base images for occupations like pilot are often Caucasian-presenting, reflecting Stable Diffusion bias.
Experimental setup
Datasets: VQAv2 FairFace UTKFace
Benchmarks: occupations dataset (diffusion perturbations)
Metrics: fairness metric (1 minus median std of true-label probabilit classification accuracy realism score race fidelity score
Baselines: FLAVA CLIP-OpenAI CLIP-LAION400M CLIP-LAION2B LENS
Builds on: Stable Diffusion Stable Diffusion XL Stable Diffusion inpainting FairFace SDEdit
Closest papers by meaning
- Finetuning Text-to-Image Diffusion Models for Fairness (ICLR 2024 · similarity 0.91)
- FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.87)
- Classes Are Not Equal: An Empirical Study on Image Recognition Fairness (CVPR 2024 · similarity 0.87)
- SocialCounterfactuals: Probing and Mitigating Intersectional Social Biases in Vision-Language Models with Counterfactual Examples (CVPR 2024 · similarity 0.87)
- Fair Text-to-Image Diffusion via Fair Mapping (AAAI 2025 · similarity 0.86)
- Balancing Act: Distribution-Guided Debiasing in Diffusion Models (CVPR 2024 · similarity 0.86)
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.