Marc S. Walton and Astrid Harth present an AI framework for automating stylistic analysis while preserving the descriptive and evidentiary concerns of art history.

A vision transformer encodes paintings and metadata. Sparse dictionary learning identifies recurring features. A large language model interprets those features through associated artworks and curator-written texts. Finally, an autonomous coordinator tests and refines the evidence into cohesive descriptions or comparisons.

Submitted 30 August 2026. Subjects: Computer Vision and Pattern Recognition, Artificial Intelligence, and Computation and Language.