Every painting is passed through a CLIP image model (ViT-B/32, open-source) that turns the picture into a 512-number vector capturing what it looks like (composition, palette, subject, handling), with no captions or labels involved. Two works are “similar” when their vectors point the same way (cosine similarity); that drives Visually closest and Painters nearest this work in each painting’s card.

This map lays all embedded works out in 2-D with t-SNE, so visually alike paintings land near each other. The dots can be coloured by painter, period or school, or by each work’s own dominant colour (the average of its pixels), captured while embedding. Painters nearest this work and the “If you like … try” chips average a painter’s vectors into one “fingerprint” and compare fingerprints.

Everything is computed offline and shipped as small precomputed files: the page itself runs no AI. Coverage grows as more works are embedded. Images © their sources via Wikimedia Commons.