Convert the images to embeddings, and perform an approximate nearest neighbor search on them and identify images that are very close together (e.g. with faiss,

Convert the images to embeddings, and perform an approximate nearest neighbor search on them and identify images that are very close together (e.g. with faiss, which the page alludes to using). It's performant enough even at scale.