Research map / Image generation / Text-to-image diffusion

OASIS Uncovers: High-Quality T2I Models, Same Old Stereotypes

ICLR 2025 · Image generation › Text-to-image diffusion · approach: Fairness and bias in generative models · 0 citations (OpenAlex)

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 OASIS, a sociologically grounded toolbox (Stereotype Score, WALS, StOP, SPI) to measure and trace stereotypes in text-to-image models, and uses it to show newer T2I models still carry strong stereotypes.

Problem

Existing stereotype measures for text-to-image models rely on statistical parity, which treats any distributional bias as a stereotype and ignores directionality, so they misclassify biases as stereotypes and offer no account of where stereotypes originate inside the model.

Approach

Defines stereotype as directional over-representation of an attribute relative to real-world statistics (max(0, P(A|D,C) - P*(A|C)) >= margin), and pairs it with a spectral variance metric (WALS) and latent-space analyses (StOP for discovering internal associations, SPI for tracking attribute emergence during denoising).

Key results

• FLUX.1 depicts 83.6% of Iranian images as men versus 98% (SDv2) and 99.8% (SDv3), yet still shows high stereotype scores such as 83.5% beard for Iranians.
• SDv3 has lower WALS across nationalities, showing that lower stereotype scores can come at the cost of lower attribute variance.
• Stereotype scores decrease as the nationality's Internet user population increases, across Iranian, Mexican, and Indian groups and all three models.

Measured effects reported in the paper

Surprising findings

• Newer models such as FLUX.1 and SDv3 show lower stereotype scores overall but still exhibit strong stereotypes, and FLUX.1 shows higher mustache stereotypes for Mexican people than SDv2 and SDv3.
• Gender imbalance for doctors worsens when a nationality is added (e.g., Iranian doctor is 100% man in SDv3 versus 78% for doctor alone).

Limitations

• Relies on CLIP zero-shot attribute classifiers, which can be inaccurate for some attributes (e.g., 83% accuracy for beard on CelebA) and may be biased.
• Requires real-world reference distributions P*(A|C) which are unavailable for many attributes; some assigned a 50% presence prior.
• Evaluation covers only three nationalities and three open-weight T2I models; StOP cluster identification is manual.

Experimental setup

Datasets: LAION2B CelebA US doctor demographic statistics (Brotherton & Etzel 2023) census and online statistics for nationality attributes

Benchmarks: CelebA

Metrics: stereotype score WALS SPI classifier accuracy

Baselines: statistical parity bias metric (Eq. 1) Jha et al. 2024 stereotype tendency

Builds on: ZeroCap D'Incà et al. 2024 open-set stereotype candidates supervised principal component analysis CLIP

Closest papers by meaning

• Association of Objects May Engender Stereotypes: Mitigating Association-Engendered Stereotypes in Text-to-Image Generation (NeurIPS 2024 · similarity 0.89)
• SocialCounterfactuals: Probing and Mitigating Intersectional Social Biases in Vision-Language Models with Counterfactual Examples (CVPR 2024 · similarity 0.87)
• Implicit Bias Injection Attacks against Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.86)
• Beyond Aesthetics: Cultural Competence in Text-to-Image Models (NeurIPS 2024 · similarity 0.86)
• Multi-Group Proportional Representations for Text-to-Image Models (CVPR 2025 · similarity 0.86)
• StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMs (NeurIPS 2026 · similarity 0.85)