This repository contains the code and data for our paper:
When Cultures Meet: Multicultural Text-to-Image Generation Parth Bhalerao, Oana Ignat, Brian Trinh, Mounika Yalamarty ACL Findings 2026
We introduce multicultural text-to-image generation as a new task, where a person from one cultural background is depicted in the context of a landmark from a different culture. We present MosAIG, a Multi-Agent framework for Image Generation that leverages LLMs with distinct cultural personas to generate richer, more culturally grounded image captions.
Our benchmark contains 9,000 images spanning:
- 5 countries (United States, Germany, India, Spain, Vietnam)
- 5 languages (English, Hindi, German, Spanish, Vietnamese)
- 3 age groups (Child, Adult, Elder)
- 2 genders (Male, Female)
- 25 historical landmarks
MosAIG/
├── Alt/ # AltDiffusion image generation pipeline
│ ├── AltDiffusion-m18-Running-Instructions.pdf
│ ├── BatchImageGenerationAltDiffusion.py
│ ├── PromptTranslation.py
│ ├── RunAltExtended.py
│ ├── main-2.py
│ └── main-3.py
├── Flux/ # FLUX image generation pipeline
│ ├── BatchImageGenerationFlux.py
│ └── FluxNotebook.zip
├── Metrics-Code/ # Evaluation metrics
│ ├── Metrics-1.py
│ ├── Metrics-2.py
│ ├── Metrics-3.ipynb
│ └── Metrics-4.ipynb
├── Multi-Agent-Setup/ # MosAIG multi-agent framework
│ ├── Final-Multi-V2.py
│ └── Simple-Crew-Setup.py
├── Single Agent Hardcoded Prompts/ # Simple baseline
│ └── SingleHardcoded-Prompts.py
├── data/ # Metadata and example data
│ ├── example_data/
│ ├── Alt_MultiAgent.xlsx
│ ├── Alt_Simple.xlsx
│ ├── Flux_Multiagent.xlsx
│ └── Flux_Simple.xlsx
└── images/ # Sample generated images
The full dataset of 9,000 images along with metadata is available on HuggingFace:
🤗 AIM-SCU/Multi-Cultural-Single-Multi-Agent-Images
The data/ folder in this repository contains the metadata spreadsheets with image captions and demographic keywords for all four model settings:
Alt_MultiAgent.xlsx— AltDiffusion multi-agent captionsAlt_Simple.xlsx— AltDiffusion simple captionsFlux_Multiagent.xlsx— FLUX multi-agent captionsFlux_Simple.xlsx— FLUX simple captions
MosAIG is a model-agnostic multi-agent prompting framework that decomposes cultural and demographic reasoning across three specialized agents:
- Country Agent — describes culturally appropriate attire and accessories
- Landmark Agent — describes architectural and environmental details
- Age-Gender Agent — describes demographic-specific physical traits
These agents interact over two rounds of conversation, supervised by a Moderator Agent, and the outputs are summarized by a Summarizer Agent into a final image caption (≤77 tokens).
The framework is implemented using CrewAI and LLaMA 3.1-8B.
We evaluate four model configurations:
| Model | Description |
|---|---|
| Alt-S | AltDiffusion + simple prompt |
| Alt-M | AltDiffusion + multi-agent prompt |
| Flux-S | FLUX + simple prompt |
| Flux-M | FLUX + multi-agent prompt |
- AltDiffusion (
BAAI/AltDiffusion-m18): multilingual open-source T2I model supporting 18 languages - FLUX (
Freepik/flux.1-lite-8B-alpha): state-of-the-art English-language T2I model
We evaluate across five dimensions:
| Metric | Description |
|---|---|
| Alignment | CLIPScore — text-image semantic correspondence |
| Quality | Inception Score — visual fidelity and diversity |
| Aesthetics | SigLIP-based aesthetic predictor (1–10 scale) |
| Knowledge | Sensitivity to landmark identity via caption swapping |
| Fairness | Consistency of performance across demographic substitutions |
Install the required packages:
pip install crewai
pip install diffusers transformers accelerate
pip install torch torchvision
pip install openai-clip
pip install pandas openpyxl
If you use this work, 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" }
- 📄 Paper: ACL Findings 2026
- 🤗 Dataset: HuggingFace
- 💻 Code: GitHub