Search by what is in the picture
Type it: cows, birthday cake,
receipt. Every photo is turned into a vector by
SigLIP 2, a model trained to put an image and a
sentence describing it close together, and a search is the
photos nearest to what you typed. It reads German as well as
English: kühe auf der weide works.
It runs on your CPU through ONNX Runtime: about 0.1 s per photo to index on a laptop, and a few tens of milliseconds per query. Results come by relevance; one switch orders the same matches by date.