Immich vs Paperless-ngx

Both of these are "set it and forget it" apps, and both earn their place on a small machine for the same reason: the expensive work happens once, at import, and afterwards you only pay for the occasional new file.

That is also where they differ. Immich's first pass is heavy and continuous while it lasts — thumbnails, face detection and OCR, all on your CPU. Paperless-ngx processes one page at a time, asynchronously, and is done in seconds.

Pick Immich if

• Your problem is photos: tens of thousands of them, scattered, and no way to find anything.
• You want to search by description rather than filename. CLIP-backed search for "beach at sunset" is the feature that justifies the install.
• You have a real photo library and are willing to give up one evening of CPU. On our 12,241-file library the crawl itself was about 5 seconds; thumbnails ran at roughly 91 photos a minute, OCR at about 18, so a 12,000-file library is around two hours of solid CPU on two cores.
• You can give it 4 GB. That is the floor with video transcoding disabled, and transcoding should stay disabled on hardware like this.

The cost is administrative as much as computational: Immich is Postgres plus Redis plus the server, three containers minimum, and the machine-learning container is the single biggest CPU consumer on a small NAS. You can move that container to an idle laptop over the network and leave the server where it is — the best trick on this list.

Pick Paperless-ngx if

• Your problem is documents: invoices, receipts, contracts, warranty PDFs, and a scanner that produces files nobody reads.
• You want full-text search over paper, not over pictures.
• You have 2 GB and nothing else to spare. It is the lightest genuinely useful tool here: the app plus a single Postgres, no GPU, no model, no cold-start penalty.
• You want it to learn. After roughly a hundred invoices the classifier starts suggesting correspondents and document types on its own.

Be clear about what the "AI" is: Tesseract OCR plus a trained classifier. It reads your scans and files them. It does not summarise them, and it will not answer questions about them. For anything but English, install the Tesseract language packs or the OCR output is noise.

The number that decides it

The storage result applies to both and is the cheapest upgrade you can make: the same interface went from 24 seconds to 0.25 seconds for its first paint when its database and thumbnails moved from a 5400 rpm 2.5" drive to a 256 GB SATA SSD. Photo and document blobs are fine on spinning rust; the database is not.

One shared trap: do not start both importers at the same moment. Two indexing jobs on a two-core machine do not run twice as fast, they run at half speed each and make the NAS unresponsive in the evening. Copy the files over first — at 11.4 MB/s on a 100 Mbps link the transfer is usually the longer half of the job — then let one indexer run to completion.

Verdict

If you have photos you cannot find, install Immich and accept the two hours. If you have paper you cannot find, install Paperless-ngx and be surprised at how little it costs.

They coexist comfortably on 8 GB, because their expensive phases rarely overlap: Immich at import, Paperless on each new scan. On 4 GB, pick one. Immich is the one that needs the RAM, so it is the one that gets the machine.