Why Image Search Returns Dogs for "No Dogs"
Summary
Search an image index for "white dog" and the first results are right, then white cars and white walls appear. Search "no dogs" and you get dogs. CLIP-style embeddings bind attributes to objects weakly (ARO, ICLR 2023) and often score at chance on negated queries (NegBench, 2025). The fix needs no retraining: recall about 100 candidates with the embedding on the positive part of the query, have a vision model check the exclusions and attributes on each image, then diversify with MMR. Store clicks and labels, cluster the queries that fail, and you can learn a small query correction without re-indexing. Explore: mixpeek.com/guides/visual-search-negation-attribute-bleed-unseen-classes
About this video
Search an image index for "white dog" and the first results are right, then white cars and white walls appear. Search "no dogs" and you get dogs. CLIP-style embeddings bind attributes to objects weakly (ARO, ICLR 2023) and often score at chance on negated queries (NegBench, 2025). The fix needs no retraining: recall about 100 candidates with the embedding on the positive part of the query, have a vision model check the exclusions and attributes on each image, then diversify with MMR. Store clicks and labels, cluster the queries that fail, and you can learn a small query correction without re-indexing. Explore: mixpeek.com/guides/visual-search-negation-attribute-bleed-unseen-classes