Guide
Films, music and photographs served to every screen in the house — and the one hardware question that decides how much server you need.
The hardware question for a media library has one name: transcoding. When a client can play a file as it is stored — the right container, the right codec, a resolution the screen accepts — the server does almost no work, and a small machine serves several people at once. When it cannot, the server re-encodes the stream in real time, and that is a different class of computer: several cores kept busy per stream, or a GPU doing it properly.
Most of the fix is in the library rather than the server. Files that match what your televisions and phones can decode play directly; one unusual codec or a subtitle format that has to be burned in turns every playback into a re-encode. It is worth checking what your clients actually support before concluding that the server is too slow.
Storage grows differently for each half. A film collection is large from the first day and then stable; a photo library starts small and grows every week, with thumbnails and machine-learning artefacts on top of the originals. Plan the disk for photographs and the processor for video, and the rest of the set is comfortable on modest hardware.
Sets
The sum of the minimums the recipes below declare. It is a floor, not a recommendation: leave room for the operating system, backups and growth.
Why it is here: The centre of a media library: film and television for every screen, with no account and no relay service in the middle. Direct play costs almost nothing; a transcode is the entire hardware question below.
Why it is here: Music is the easy half. Files stream as they are, dozens of Subsonic clients already exist, and the whole thing runs in 256 MB — the smallest useful service in this catalog.
Why it is here: Photographs and video from phones, with backup, deduplication and search that actually finds things. It expects real hardware and says so: 6 GB of RAM and 20 GB of disk before your library arrives.
Why it is here: A photo library for a collection that already exists on a disk, with local machine learning for search and faces. Lighter than Immich, and less focused on continuous phone backup.