Skip to main navigation Skip to search Skip to main content

Towards Culturally Aware AI: Quantified Evaluation of Relevance and Similarity in AI-Generated Images

  • Hong Kong University of Science and Technology

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Despite advances in text-to-image models, issues like cultural bias and inaccuracy persist, underscoring the need for reliable quantitative metrics to evaluate cultural representation. This work introduces CRIX, a prompt-guided, context-aware Cultural Relevance Index that leverages visual language models (VLMs) and visual question answering (VQA) to assess both the content and context of AI-generated images. Applied to Arabic and South Asian cultures, CRIX achieves lower mean squared errors (0.0022 and 0.0053, respectively) than baseline metrics, indicating stronger alignment with human evaluations. We also propose the Cultural Similarity Score (CSS) to quantify how similarly two images reflect a target culture.

Original languageEnglish
Title of host publicationAdvances In Computer Graphics, Cgi 2025, Pt Iii
EditorsP Li, L Ma, L Wan, B Sheng, J Kim, D Thalmann, N Magnenat-Thalmann
PublisherSpringer Science and Business Media Deutschland GmbH
Pages331-340
Number of pages10
Volume16509
ISBN (Electronic)978-3-032-22267-1
ISBN (Print)9783032222664
DOIs
Publication statusPublished - 2 Jul 2026
Event42nd Computer Graphics International Conference, CGI 2025 - Hong Kong, China
Duration: 14 Jul 202518 Jul 2025

Publication series

NameLecture Notes In Computer Science

Conference

Conference42nd Computer Graphics International Conference, CGI 2025
Country/TerritoryChina
CityHong Kong
Period14/07/2518/07/25

Keywords

  • AI-Generated Content
  • Evaluation Methods
  • Generative Models

Fingerprint

Dive into the research topics of 'Towards Culturally Aware AI: Quantified Evaluation of Relevance and Similarity in AI-Generated Images'. Together they form a unique fingerprint.

Cite this