Immich

Why it's ranked here

The engineering is unusually serious for a self-hosted project. Business logic sits in about fifty services kept separate from storage and process adapters, boot waits on a Postgres advisory lock and can come up in maintenance mode, and search quality is published as recall, memory and latency figures per model instead of adjectives. The cost is operational: four containers, a pinned database image, and a data directory you must back up yourself. Copyleft licensing under AGPL v3 also rules it out as a component inside a closed product. For a home library you actually care about, this is the strong default.

What's good

Ownership is explicit. Originals are never rewritten, and edited metadata goes to companion sidecar files and the database. The server takes its own database dumps on a schedule, keeping the last fourteen by default, and a restore writes a rollback point first, then health checks afterward. Search reaches faces, text found inside images, folder paths, camera model, star rating and freeform description, with multilingual models for people who do not type queries in English. Hardware acceleration is a build argument for both transcoding and inference, so one setup covers a small always-on box or a GPU host.

Tradeoffs

This is not a light install. The stock deployment runs a server, an inference container, a pinned Postgres image carrying a vector extension, and a queue, plus optional metrics and dashboards. Database dumps hold metadata only, and the project states plainly that it does not back up your files for you, so a real backup plan stays your job. Vision models are large: the documented English models range from roughly 1,000 to 6,600 MiB resident. Self-signed certificates, basic auth and mutual TLS on mobile are labelled experimental and not actively developed. There is no filtering of sensitive imagery.

How to use it well

Best fit is a household replacing a cloud photo service: one machine you control, a few accounts, partner sharing, phones backing up automatically. Budget memory for the inference container, pick the search model to match the languages your users type in, and reprocess assets after any model change. Pair it with a real backup tool and offsite copies, because the built-in dumps cover metadata only. It does not try to be a general file sync tool, a document archive, or an editing suite, and the licence makes it a poor building block inside a proprietary product.

Observed

License

AGPL v3

Server

TypeScript on Node, NestJS with the Kysely query builder

Inference service

Python and FastAPI, models loaded in ONNX format

Data stores

Postgres with a vector index extension, plus a Redis compatible queue

Clients

Flutter mobile app, SvelteKit web app, npm command line uploader

Install surface

Docker Compose, four core containers plus optional metrics and dashboard containers

Interfaces

REST API with OpenAPI-generated clients, API keys, OAuth login, plugin SDK package

Acceleration

Build-time device selection for transcoding and for inference, including CUDA, ROCm, OpenVINO, RKNN and ArmNN

Repo layout

pnpm monorepo covering server, web, mobile, machine learning, docs and packages