CIFAR-100 — image scatter

50k CLIP images on a deck.gl map — every image at its UMAP coord, outlined by cluster. Compare dyf vs k-means at matched granularity.

Most of the gallery shows clusterings as colored dots. This one shows the actual images: 10,000 CIFAR-100 thumbnails (from 50k CLIP embeddings) placed at their UMAP coordinates, each backed by a square outline colored by its cluster.

Use the toolbar to interrogate the structure:

• granularity (k) — re-cut the clustering from 2 to 128 clusters.
• dyf / k-means — recolor outlines by either method at the same k. dyf gives nested tree cuts (one fit → every k); k-means is an independent fit per k.
• images on/off — hide the photos to read the cluster coloring cleanly.

Honest result: against the 20 true coarse classes both methods land around NMI ≈ 0.45 — dyf edges k-means at coarse k (≤16), k-means edges dyf at fine k (≥32). CIFAR-100 does not cluster trivially in CLIP space. Build it yourself with demo/cifar_deck_demo.py.