CultiVATE Benchmark

CultiVATE benchamrk was used in Culture in Action: Evaluating Text-to-Image Models through Social Activities (ICLR 2026) for evaluating T2I generative models on cultural activities. This page includes the benchmark datasets, please visit the github repo in order to use the AHEaD automatic metrics.

• Paper link
• Project page
• Code

CultiVATE covers 576 social activities in 9 categories (for example dancing, greeting, and dining) across 16 countries. This dataset contains the generated images: about 19,000 images from six T2I models.

Each row is one generated image.

column	meaning
image	generated PNG
prompt	text-to-image prompt
activity	e.g. dancing
subactivity	e.g. line dance
country	e.g. USA
model	FLUX.1-dev, stable-diffusion-3.5-medium, qwen-image, dall-e-3, gpt-image-1, or gemini-2.5-flash-image-preview
image_index	1–10 for open models; 1 for proprietary models

Prompt template: A photorealistic photo of {phrase} in {country}.

from datasets import load_dataset

ds = load_dataset("sinamalakouti/CultiVATE", "generated", split="test")
print(ds[0]["prompt"], ds[0]["activity"], ds[0]["subactivity"], ds[0]["country"])

Real images used as references for the EXAG (exaggeration) metric are in sinamalakouti/CultiVATE-real (access requires approval). They are not included here.

Citation

@inproceedings{ICLR2026_dca63f26,
author = {Malakouti, Sina and Gong, Boqing and Kovashka, Adriana},
booktitle = {International Conference on Learning Representations},
editor = {C. Vondrick and B. Hariharan and C. Raffel and L. Pinto and D. Yang and A. Faust},
pages = {136235--136262},
title = {Culture in Action: Evaluating Text-to-Image Models through Social Activities},
url = {https://proceedings.iclr.cc/paper_files/paper/2026/file/dca63f2650fe9e88956c1b68440b8ee9-Paper-Conference.pdf},
volume = {2026},
year = {2026}
}