EMNLP Findings2023
'Person' == Light-skinned, Western Man, and Sexualization of Women of Color: Stereotypes in Stable Diffusion
Sourojit Ghosh, Aylin Caliskan
Abstract
We study stereotypes embedded within one of the most popular text-to-image generators: Stable Diffusion. We examine what stereotypes of gender and nationality/continental identity does Stable Diffusion display in the absence of such information i.e. what gender and nationality/continental identity is assigned to 'a person', or to 'a person from Asia'. Using vision-language model CLIP's cosine similarity to compare images generated by CLIP-based Stable Diffusion v2.1 verified by manual examination, we chronicle results from 136 prompts (50 results/prompt) of front-facing images of persons from 6 different continents, 27 nationalities and 3 genders. We observe how Stable Diffusion outputs of 'a person' without any additional gender/nationality information correspond closest to images of men (avg. cosine similarity 0.64) and least with persons of nonbinary gender (avg. cosine similarity 0.41), and to persons from Europe/North America (avg. cosine similarities 0.71 and 0.68, respectively) over Africa/Asia (avg. cosine similarities 0.43 and 0.41, respectively), pointing towards Stable Diffusion having a concerning representation of personhood to be a European/North American man. We also show continental stereotypes and resultant harms e.g. a person from Oceania is deemed to be Australian/New Zealander (avg. cosine similarities 0.77 and 0.74, respectively) over Papua New Guinean (avg. cosine similarity 0.31), pointing to the erasure of Indigenous Oceanic peoples, who form a majority over descendants of colonizers both in Papua New Guinea and in Oceania overall. Finally, we unexpectedly observe a pattern of oversexualization of women, specifically Latin American, Mexican, Indian and Egyptian women relative to other nationalities, measured through an NSFW detector. This demonstrates how Stable Diffusion perpetuates Western fetishization of women of color through objectification in media, which if left unchecked will amplify this stereotypical representation. Image datasets are made publicly available.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
11 top-tier papers have cited it since. Know what they found before your reviewers do.
Lune can draw on
the full paper · the top-tier papers that cite it · related work
Free to start. No credit card required.
Lineage
- ViSAGe: A Global-Scale Analysis of Visual Stereotypes in Text-to-Image Generation cites High-Resolution Image Synthesis with Latent Diffusion Models
- ViSAGe: A Global-Scale Analysis of Visual Stereotypes in Text-to-Image Generation cites T2IAT: Measuring Valence and Stereotypical Biases in Text-to-Image Generation
- POET: Supporting Prompting Creativity and Personalization with Automated Expansion of Text-to-Image Generation cites High-Resolution Image Synthesis with Latent Diffusion Models
- T2IAT: Measuring Valence and Stereotypical Biases in Text-to-Image Generation cites High-Resolution Image Synthesis with Latent Diffusion Models
Cited by top-tier papers11
- Generative AI and Perceptual Harms: Who's Suspected of using LLMs?Kowe Kadoma, Danaë Metaxa, Mor NaamanCHI 2025 · 28 citations
- ViSAGe: A Global-Scale Analysis of Visual Stereotypes in Text-to-Image GenerationAkshita Jha, Vinodkumar Prabhakaran, Remi Denton, Sarah Laszlo et al.ACL 2024 · 18 citations
- VisBias: Measuring Explicit and Implicit Social Biases in Vision Language ModelsJen-Tse Huang, Jiantong Qin, Jianping Zhang, Youliang Yuan et al.EMNLP 2025 · 13 citations
- "We're not all construction workers": Algorithmic Compression of Latinidad on TikTokNina Lutz, Cecilia R. AragonCSCW 2024 · 9 citations
- POET: Supporting Prompting Creativity and Personalization with Automated Expansion of Text-to-Image GenerationEvans Xu Han, Alice Qian Zhang, Haiyi Zhu, Hong Shen et al.UIST 2025 · 5 citations
Builds on3
- Language Models are Few-Shot LearnersTom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah et al.NeurIPS 2020 · 64,255 citations
- High-Resolution Image Synthesis with Latent Diffusion ModelsRobin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser et al.CVPR 2022 · 13,123 citations
- T2IAT: Measuring Valence and Stereotypical Biases in Text-to-Image GenerationJialu Wang, Xinyue Gabby Liu, Zonglin Di, Yang Liu et al.ACL Findings 2023 · 26 citations
Related papers
- Do Existing Testing Tools Really Uncover Gender Bias in Text-to-Image Models?Yunbo Lyu, Zhou Yang, Yuqing Niu, Jing Jiang et al.ACM MM 2025 · 2 citations
- Partiality and Misconception: Investigating Cultural Representativeness in Text-to-Image ModelsLili Zhang, Xi Liao, Zaijia Yang, Baihang Gao et al.CHI 2024 · 17 citations
- Image-Perfect Imperfections: Safety, Bias, and Authenticity in the Shadow of Text-To-Image Model EvolutionYixin Wu, Yun Shen, Michael Backes, Yang ZhangCCS 2024 · 3 citations
- SustainDiffusion: Optimising the Social and Environmental Sustainability of Stable Diffusion ModelsGiordano d’Aloisio, Tosin Fadahunsi, Jay Choy, Rebecca Moussa et al.ICSE 2026
- Person-Centric Annotations of LAION-400M: Auditing Bias and Its Transfer to ModelsLeander Girrbach, Stephan Alaniz, Genevieve Smith, Trevor Darrell et al.ICLR 2026 · 5 citations
