A neural network (CLIP) reads every photo into a 512-dimension embedding, and UMAP folds those down to this 2D map, so visually similar photos cluster together.
