Reproduce “CuRe: Cultural Gaps in the Long Tail of Text-to-Image Systems” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aniket Rege, Zinnia Nie, Mahesh Ramesh, Unmesh Raskar, Zhuoran Yu, Aditya Kusupati, Yong Jae Lee, Ramya Korlakai Vinayak
Researchers developed CuRe, a framework that quantifies cultural representational gaps in text-to-image systems using a structured dataset and Marginal Information Attribution. Their evaluation revealed systematic disparities, with Global South artifacts showing consistently poorer cultural understanding compared to Global North counterparts across state-of-the-art models.
Key insights
- Increasing attribute specification in prompts through Marginal Information Attribution (MIA) can reveal a T2I system's true cultural understanding, indicating whether it relies on superficial pattern matching or deeper contextual knowledge.
- A
- democratic scaling
- approach, leveraging Wikimedia knowledge graphs, enables systematic and scalable construction of culturally diverse datasets, reducing the cost and bias of manual curation.
- Automated metrics for perceptual similarity, image-text alignment, and diversity can effectively correlate with human judgments for cultural representativeness and identify specific cultural knowledge gaps, particularly in the
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.