Evidence Hub

Visual Distances among Cultural Art Corpora Align with Expert-Coded Historical-Artistic Relatedness

Wang Q, Ying W, Xu Z, Yao Y, Wen S, Wu G, Yao C, Ying F et al.

2026

4,500 participants

Abstract

Abstract Visual distances derived from large vision models correlate with expert-coded historical-artistic relatedness across 53 cultural units (five continents, five millennia, 14 families). Ten blinded experts coded 1,378 pairwise judgments (Cohen's κ = 0.949, 0.931–0.972). We compared six primary models via Gromov-Wasserstein distance, validating with Quadratic Assignment Procedure (10,000 permutations), Bayesian inference, and geographic controls. A supplementary perceptual discriminability check based on the final analyzed sample of n = 4,500 participants across 12 countries confirmed that the broad stimulus categories are visually distinct to lay observers. All six primary models showed significant negative associations between latent-space distance and expert relatedness (Spearman ρ=−0.443 to − 0.497; FDR q 60), and 100% leave-one-unit-out sign retention. Cross-model median intercorrelation was ρ=+0.446; negative controls yielded 0/10,000 permutations at observed magnitude. The signal remained detectable in exclusively transoceanic pairs. The convergence across six primary architectures indicates that internet-scale visual training corpora contain statistical regularities that correlate significantly with expert-judged art-historical relationships across cultures and epochs, though geographic proximity, training data curation, and Western art-historical frameworks cannot be fully excluded as contributing factors.

Read the paper

Related studies

Go deeper in Evidence Hub

Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.

Open this study Sign up free