Smart search and faces
Both run on a separate, optional service. Without it, everything else works.
Turning it on
Add these to .env:
COMPOSE_PROFILES=ml
ML_URL=http://ml:8000
Then apply it:
docker compose up -d
The models download on first use, about 1 GB, and the service needs about 2 GB of memory. Progress is under Admin → Settings. Pictures you already have are analysed in the background.
Searching
Describe what you are looking for, in your own language. You can also filter by date, person, place and camera, and words inside pictures are found too.
People
Faces are grouped into people, and you can name, merge and hide them. Names and faces stay private: partners, album members and guests do not see them, and nothing about faces leaves your server.
ML_FACE_MODEL chooses the model: buffalo_l by default, or buffalo_s for weaker hardware.
The default face models are licensed for non-commercial use only. Check their terms if you use Mediarr commercially.