Exposing Blindspots: Cultural Bias Evaluation in Generative Image Models

Huichan Seo, Sieun Choi, Minki Hong, Yi Zhou, Junseo Kim, Lukman Ismaila, Naome Etori, Mehul Agarwal, Zhixuan Liu, Jihie Kim, Jean OharXiv 2025 · DOI · arXiv

Introduction

Generative AI models have reached a level of visual sophistication that often masks underlying cultural biases. While text-to-image (T2I) generation has been widely studied, the iterative process of image-to-image (I2I) editing—where users attempt to refine or "fix" outputs—remains a black box regarding cultural fidelity. This research introduces a framework to audit these models, revealing a persistent "Global-North" default and a phenomenon termed "metric drift," where automated scores fail to capture the erosion of cultural authenticity.

The ECB Corpus and Evaluation Framework

To systematically evaluate cultural bias, the researchers developed the Exposing Cultural Blindspots (ECB) corpus. This dataset spans six countries—China, India, Kenya, Korea, Nigeria, and the U.S.—and covers 36 cultural subcategories across eight top-level domains such as Architecture, Food, and Art.

Overview of cultural categories and country-specific generations in the ECB corpus.

The evaluation pipeline utilizes era-aware prompting (traditional vs. modern) and multi-step I2I editing loops. To move beyond standard metrics, the authors implemented a Retrieval-Augmented Generation (RAG) based metric that verifies specific cultural markers against Wikipedia data, validated by a panel of 17 domain experts.

The "Global-North" Default and Metric Drift

A primary finding is the models' tendency to default to Western aesthetics when prompts are country-agnostic. Even when specific countries are mentioned, the models often produce "Westernized" versions of local traditions, failing to adapt phenotypes or architectural styles appropriately.

Examples of the "Global-North" default in India, where a "traditional wedding" and "house" reflect Western modern styles.

More critically, the study identifies Metric Drift. During iterative editing, standard metrics like CLIPScore often remain stable or improve, suggesting the model is successfully following the text prompt. However, human expert audits reveal a sharp decline in cultural quality—averaging a 44.2% drop in Human Quality Score (HQS) by the fifth edit. The models often resort to "sticker" editing, adding superficial cues like flags rather than deep, contextually accurate cultural transformations.

Results and Implications

The audit of five major model families, including FLUX.1 and Stable Diffusion 3.5, shows that while some models have higher initial alignment, all suffer from cultural decay during editing. Nigeria and the U.S. were found to be particularly vulnerable to quality loss in repetitive loops. Furthermore, occupational audits confirmed that deep-seated demographic biases persist, with high-status roles predominantly rendered as light-skinned males.

This work highlights the urgent need for generative models that move beyond superficial cultural markers toward deep, context-preserving representations. Future developments should focus on finer-grained cultural data and training objectives that penalize "shortcut" cues in favor of authentic diversity.

Summary updated

Abstract

Generative image models produce striking visuals yet often misrepresent culture. Prior work has examined cultural bias mainly in text-to-image (T2I) systems, leaving image-to-image (I2I) editors underexplored. We bridge this gap with a unified evaluation across six countries, an 8-category/36-subcategory schema, and era-aware prompts, auditing both T2I generation and I2I editing under a standardized protocol that yields comparable diagnostics. Using open models with fixed settings, we derive cross-country, cross-era, and cross-category evaluations. Our framework combines standard automatic metrics, a culture-aware retrieval-augmented VQA, and expert human judgments collected from native reviewers. To enable reproducibility, we release the complete image corpus, prompts, and configurations. Our study reveals three findings: (1) under country-agnostic prompts, models default to Global-North, modern-leaning depictions that flatten cross-country distinctions; (2) iterative I2I editing erodes cultural fidelity even when conventional metrics remain flat or improve; and (3) I2I models apply superficial cues (palette shifts, generic props) rather than era-consistent, context-aware changes, often retaining source identity for Global-South targets. These results highlight that culture-sensitive edits remain unreliable in current systems. By releasing standardized data, prompts, and human evaluation protocols, we provide a reproducible, culture-centered benchmark for diagnosing and tracking cultural bias in generative image models. Project page: https://seochan99.github.io/ECB
BibTeX
@misc{seo2025exposing,
  author = {Huichan Seo and Sieun Choi and Minki Hong and Yi Zhou and Junseo Kim and Lukman Ismaila and Naome Etori and Mehul Agarwal and Zhixuan Liu and Jihie Kim and Jean Oh},
  title = {{Exposing Blindspots: Cultural Bias Evaluation in Generative Image Models}},
  year = {2025},
  eprint = {2510.20042},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.20042},
}
Figure