

Also, if it turns out that immich misses too many faces, or has too many false positives, or creates too many different persons for one face, or groups different persons together, you can tune that. You can set the threshold values for all these things, or just use a different ML model and reanalyse your whole library. Even on an older CPU without any special ML features, it takes less than a night to reanalyse a hundred thousand pictures or so.

It is fundamentally built around the files being decrypted in RAM, for all the search features. You can use an encrypted partition for storing the photos and DB, to avoid having plaintext files on disk.