An atlas of learned visual similarity
DINOv2 has never been told what a flag is. Given nothing but pixels, it sets the Nordic crosses side by side, keeps the British ensigns together across an ocean, and scatters communist red to the far corners of its space. This is the map it drew — and you can walk around inside it.
Sweden, Denmark, Norway, Iceland and Finland sit in the tightest cluster of any tradition — 64% k=5 nearest-neighbour purity in 384 dimensions. The structure carries the signal, not the colour: a single off-centre cross survives every palette it is painted in.
The UK, Australia, New Zealand, Fiji and Tuvalu share a canton-on-blue template, and despite wildly different fly designs the encoder files them together. Australian state flags pile in on top of them — a 31× k-NN lift over chance.
Europe scatters across the whole map. Latin America does not: the horizontal tricolour with a central charge forms one tight subcluster spanning two continents. Argentina, Uruguay, Honduras, El Salvador — visual cousins, political coincidence.
China, Vietnam, North Korea, Belarus, Mongolia, Angola — a shared colour, not a shared structure. Because the encoder reads layout over palette, a red field with a star, a hammer-and-compass, an ornate seal and a black stripe end up nowhere near each other.
Twenty-five Martian regional flags, each procedurally derived from an Earth tradition. 13 of 25 land closest to the centroid they inherited from. Structurally distinctive parents (Nordic, solid+emblem, pan-Arab) pass the resemblance down; colour-only parents (heraldic, pan-African) do not.
| Encoder | DINOv2 ViT-S/14, frozen. Self-supervised on natural images; never fine-tuned, never shown a label. |
|---|---|
| Corpus | 423 flags — 197 sovereign, 201 subdivision, 25 Martian. SVGs rasterized before encoding. |
| Alignment | Two encoder runs (398 flags at 518 px, 197 at 224 px) joined into one 384-d frame by orthogonal Procrustes on the 197 shared sovereigns. |
| Projection | 3D PCA · 3D t-SNE (perplexity 20, cosine) · 3D UMAP (n=15, min_dist=0.05). |
| Palette | Four-colour median cut per flag, classified into twelve named families. Used for the palette colouring, never for the embedding. |
| Lineage | A visual analogue of Figure 3 in emoji2vec. Sister project: music2vec. |