The same 165 product images, embedded two ways.Pick a product and each panel shows the ten nearest that model actually returns.
Pick a product and its ten nearest neighbors get drawn in both panels, with the distance each model returns. Neighbors of a different type are marked, so you can see where a model confuses one category for another. Hover any dot for its type.
Dot positions are the saved UMAP layouts from the article figure, for display only. Every distance and every neighbor list comes from the native vectors: cosine over 512 dimensions for CLIP, and the Lorentz geodesic over 513 for Hyper3-CLIP. The two layouts use different output geometries, so visual separation on its own does not establish a retrieval improvement. The neighbor counts do.