What machines see

This page is a record of how different systems have looked at this work.

The embeddings — rows of numbers generated by models like CLIP and DINOv2 — describe the image mathematically: shapes, colours, spatial relationships, visual texture. They power the similar-works connections above, finding echoes across thirty years of work without any manual tagging. Each model weighs visual features differently. CLIP learned from image-text pairs; DINOv2 learned from images alone. Whether they agree on what's similar is itself informative.

The vision analyses are the language layer — what happens when a model is asked to describe what it actually sees. The same prompt, sent to different models, on different dates. Where they converge, something is probably objectively there in the work. Where they diverge, the work is leaving room for interpretation.

Both are dated and model-tagged because they're not timeless descriptions. They're records of a specific system looking at a specific image at a specific moment. As models improve, the readings get richer. The archive accumulates them all.