AIM-SCU/When-Cultures-Meet

When Cultures Meet: Multicultural Text-to-Image Generation

This repository contains the dataset released with our paper When Cultures Meet: Multicultural Text-to-Image Generation, published in Findings of ACL 2026.

We introduce multicultural text-to-image generation, where people and landmarks from different cultures are represented together within the same generated scene.

The benchmark contains 9,000 AI-generated images spanning 5 countries, 5 languages, 3 age groups, 2 genders, and 25 historical landmarks.

Images were generated using AltDiffusion and FLUX, comparing simple prompting against our MosAIG multi-agent prompting framework.

Dataset

Subset	Model	Prompting	Images
Alt_Multi_V2	AltDiffusion	MosAIG	3,750
Alt_Single	AltDiffusion	Simple	3,750
Flux_Multi_V2	FLUX	MosAIG	750
Flux_Single	FLUX	Simple	750
Total			9,000

Each image subset has a corresponding .xlsx metadata file.

Files

• Alt_Multi_V2.xlsx — metadata for the 3,750 AltDiffusion MosAIG images
• Alt_Single.xlsx — metadata for the 3,750 AltDiffusion simple-prompt images
• Flux_Multi_V2.xlsx — metadata for the 750 FLUX MosAIG images
• Flux_Single.xlsx — metadata for the 750 FLUX simple-prompt images
• Alt_Multi_V2_3750/ — 3,750 generated PNG images
• Alt_Single_3750/ — 3,750 generated PNG images
• Flux_Multi_V2_750/ — 750 generated PNG images
• Flux_Single_750/ — 750 generated PNG images

Cultural and Landmark Coverage

The benchmark combines people from different cultural backgrounds with landmarks associated with five countries.

Germany

• Cologne Cathedral
• Reichstag Building
• Neuschwanstein Castle
• Brandenburg Gate
• Holocaust Memorial

India

• Taj Mahal
• Lotus Temple
• Gateway of India
• India Gate
• Charminar

Spain

• Sagrada Familia
• Alhambra
• Guggenheim Museum
• Roman Theater of Cartagena
• Royal Palace of Madrid

United States

• White House
• Statue of Liberty
• Mount Rushmore
• Golden Gate Bridge
• Lincoln Memorial

Vietnam

• Meridian Gate of Huế
• Independence Palace
• One Pillar Pagoda
• Ho Chi Minh Mausoleum
• Thien Mu Pagoda

The benchmark systematically varies cultural, demographic, landmark, and linguistic attributes to study how text-to-image generation models behave when multiple cultural identities are represented within the same scene.

MosAIG

MosAIG is the multi-agent framework explored in the accompanying paper for improving multicultural text-to-image generation.

Instead of relying only on a simple prompt, MosAIG uses multiple specialized agents to compose culturally and demographically detailed image descriptions.

The framework includes:

• Moderator Agent
• Country Agent
• Landmark Agent
• Age-Gender Agent
• Summarizer Agent

These agents collaboratively construct culturally grounded prompts before the final prompt is provided to the text-to-image generation model.

The dataset contains images generated using both the MosAIG multi-agent approach and simple prompting, enabling direct comparison between the two strategies.

Evaluation Dimensions

The accompanying research analyzes multicultural text-to-image generation across several dimensions, including:

• Alignment — correspondence between the intended prompt and generated image
• Image Quality — overall visual quality and photorealism
• Aesthetics — visual appeal of the generated images
• Knowledge — representation of landmark and cultural knowledge
• Fairness — differences in performance across demographic, cultural, and linguistic groups

Intended Use

This dataset is intended primarily for research involving:

• Multicultural text-to-image generation
• Multilingual image generation
• Cultural representation in generative AI
• Demographic and intersectional evaluation
• Cultural grounding
• Text-image alignment
• Fairness and bias analysis
• Multi-agent prompting
• Evaluation of text-to-image generation models

Limitations and Responsible Use

All images contained in this dataset are AI-generated.

They should not be interpreted as authoritative or complete representations of cultures, nationalities, demographic groups, landmarks, or cultural identities.

Text-to-image generation models may produce:

• Cultural inaccuracies
• Demographic stereotypes
• Landmark inaccuracies
• Visual artifacts
• Uneven representation across cultural groups
• Uneven performance across languages
• Uneven performance across demographic groups

These limitations are also important aspects of the research questions studied using this benchmark.

Users should therefore exercise appropriate caution when using the dataset for applications involving cultural or demographic representation.

Paper

When Cultures Meet: Multicultural Text-to-Image Generation

Authors: Parth Bhalerao, Oana Ignat, Brian Trinh, Mounika Yalamarty

Venue: Findings of the Association for Computational Linguistics: ACL 2026

Pages: 35808–35828

Location: San Diego, California, United States

Paper: https://aclanthology.org/2026.findings-acl.1783/

DOI: https://doi.org/10.18653/v1/2026.findings-acl.1783

Code: https://github.com/AIM-SCU/MosAIG

Citation

If you use this dataset, benchmark, or the MosAIG framework in your research, please cite:

@inproceedings{bhalerao-etal-2026-cultures,
    title = "When Cultures Meet: Multicultural Text-to-Image Generation",
    author = "Bhalerao, Parth and Ignat, Oana and Trinh, Brian and Yalamarty, Mounika",
    editor = "Liakata, Maria and Moreira, Viviane P. and Zhang, Jiajun and Jurgens, David",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-acl.1783/",
    doi = "10.18653/v1/2026.findings-acl.1783",
    pages = "35808--35828",
    ISBN = "979-8-89176-395-1",
    abstract = "Text-to-image generation models have achieved strong performance in culturally homogeneous settings, yet their ability to generate multicultural scenes{---}where people and landmarks originate from different cultures{---}remains largely unexplored. We introduce multicultural text-to-image generation as a new task and present the first benchmark designed to study this setting. Our dataset contains 9,000 images spanning five countries, three age groups, two genders, 25 historical landmarks, and five languages. Using this benchmark, we analyze the behavior of state-of-the-art text-to-image models across multiple dimensions, including alignment, image quality, aesthetics, knowledge, and fairness. As one strategy for composing cultural and demographic information, we explore MosAIG, a Multi-Agent framework that enhances multicultural image generation by leveraging large language models with distinct cultural personas. Our analysis shows that richer prompt composition can improve image quality and cultural grounding compared to simple prompts, while also revealing substantial disparities across languages and demographic groups. We release our dataset and code at https\://github.com/AIM-SCU/MosAIG"
}

Authors

Parth Bhalerao, Oana Ignat, Brian Trinh, Mounika Yalamarty

Santa Clara University Santa Clara, California, USA

License

This dataset is released under the MIT License.