Bradford heritage-science paper uses facial geometry on Holbein drawings to challenge two long-standing Anne Boleyn captions

Anne Boleyn was queen of England from 1533 until her execution in 1536, yet no lifetime panel portrait commands universal agreement among curators. Art historians therefore treat Hans Holbein the Younger’s court drawing books—held by the Royal Collection Trust—as a forensic front line: preparatory chalks, relabelled inscriptions, and centuries of retouching all complicate who is pictured on any given sheet.

In May 2026 public discussion of a peer-reviewed paper in npj Heritage Science, a University of Bradford visual-computing team described feeding digitised facial contours from the Holbein album into a learned comparator that scores geometric similarity across sheets. Negative controls included unrelated sitters; positive structure came from clustering against Elizabeth I as a child and documented cousins. The Royal Collection Trust said publicly it did not commission the study and does not endorse the algorithmic attributions, while welcoming serious technical attention to the drawings themselves.

How the headline reassignment is stated

The authors’ public-facing summary boils down to two provocative pairings between traditional captions and model output.

Supporters argue the reassignment aligns with written descriptions of Anne as dark-haired with a “little neck.” Sceptics reply that costume, under-drawing, pigment loss, and 18th-century re-inscriptions make chalk studies a weak analogue to passport biometrics.

Method limits both sides already accept

Holbein’s court drawings are preparatory, not finished likenesses commissioned as legal identification. Multispectral work sometimes reveals multiple authorship layers; what a camera sees today may not be what a Tudor hand last touched.

Similarity scores report geometric distance between digitised outlines; they do not, standing alone, prove who stood for the sitting. Falsification therefore requires independent teams, convergent imaging, and explicit criteria for what would overturn each caption—not a single model run promoted without replication.

Who carries the argument in public filings

Visual-computing professor Hassan Ugail co-authored with independent historian Karen L. Davies, who has described funding archival travel through paid cleaning work from August 2024 onward. Traditional connoisseurship voices counter that centuries of cumulative eye-and-archive practice should outweigh opaque clustering unless methods and weights are fully disclosed.

The clash sits inside a wider 2020s debate: computational pattern discovery in cultural heritage versus interpretive restraint when public trust in museums is already contested.

What would actually move museum labels

Technical bulletins from the Royal Collection—multispectral stacks, pigment analysis, paper provenance chains—would supply material constraints independent of any one algorithm. Peer responses in art-history journals and conference replication on other Northern Renaissance albums would test whether the geometry is stable across digitisation pipelines.

Policy guidance on displaying algorithmic attributions beside object labels, if national heritage bodies publish it, would set display rules even when scholarship remains split. Until those hinges align, the honest read is a structured question, not a closed identification.