Immich

Immich runs on a Raspberry Pi 5, because the project supports arm64 alongside amd64. But only the 8GB and 16GB boards meet its 6GB RAM minimum with machine learning switched on. Expect face detection and smart search jobs to run slowly on a Pi, and don't plan on hardware video transcoding: Immich's docs say Raspberry Pi is currently not supported for it.

#What Immich is

Immich is a self-hosted photo and video backup app built to feel like Google Photos. You install the server with Docker Compose, point the Android or iOS app at it, and your camera roll backs up to your own disk. It's developed with backing from FUTO and licensed AGPL-3.0.

The latest release is v3.2.4 (28 September 2026). A little history helps if you're reading older guides:

• v2.0.0 (1 October 2025) was billed as the first stable release.
• v3.0.0 (2 July 2026) was the next major version. Its notes say most breaking changes are API updates that mainly affect third-party tools. It also dropped pgvecto.rs, so very old installs have to migrate to VectorChord first. Updating means setting IMMICH_VERSION=v3 in .env, then pulling and restarting.

#Immich requirements

From the official requirements page:

Windows and macOS work through Docker Desktop, but the docs "strongly discourage" non-Linux hosts. Docker inside LXC containers isn't recommended either.

Two storage rules matter more than people expect. The Postgres database, usually 1–3GB, should sit on local SSD storage and "never a network share of any kind". Thumbnails and transcoded videos add roughly 10–20% to the size of your library.

#Immich on a Raspberry Pi 5, by RAM

The Pi 5 comes with 1GB, 2GB, 4GB, 8GB or 16GB of RAM. Mapping that against Immich's own numbers:

• 16GB: comfortably above the recommendation, with room to run other services alongside.
• 8GB: meets the recommended 8GB. This is the natural Immich board.
• 4GB: below the minimum. The docs say a 4GB system can run Immich only with machine learning disabled, which means no face recognition and no smart search. The same applies to a Raspberry Pi 4 with 4GB.
• 1GB and 2GB: not viable.

The Pi 5 has four Cortex-A76 cores, which meets the recommended core count. The FAQ still warns that transcoding and machine learning "can be too heavy to run on a Raspberry Pi". The first upload is the hardest part, because every photo queues thumbnail, face-detection and smart-search jobs at once.

The FAQ suggests a few ways to make that bearable:

1. Set the job concurrency for heavy jobs to 1.
2. Drop transcoding threads to 1 or 2.
3. Switch the face model from buffalo_l to the smaller buffalo_s, then re-run face detection.
4. Run the machine-learning container on a stronger machine and keep only the server on the Pi.

For storage, the Pi 5 brief lists a PCIe 2.0 x1 interface and two USB 3.0 ports, so an NVMe or USB SSD for the database is easy to arrange. Keep that database off a NAS share.

#Does Immich need a GPU?

No. Machine learning and transcoding both run on the CPU by default. A GPU is optional and, in the docs' own words, "experimental":

• Machine learning backends: CUDA (NVIDIA, compute capability 5.2+), ROCm (AMD), OpenVINO (Intel iGPUs such as Iris Xe and Arc), ARM NN (Mali GPUs only) and RKNN (Rockchip).
• Transcoding APIs: NVENC, Quick Sync, RKMPP and VAAPI. The docs list Raspberry Pi as not supported.

Both require Linux, or Windows through WSL2.

#Features

The feature list is long. The parts people switch for are phone backup, facial recognition, smart search ("sunset at the beach"), duplicate detection, memories, partner sharing, multiple users, shared albums and public links, maps, live and motion photos, RAW support, a locked folder, OAuth login, external read-only libraries and a CLI uploader. The FAQ is firm that originals stay untouched: there's no option to compress images during backup.

On iOS, background backup depends on the operating system's scheduler, so the FAQ lists steps that make it more likely to run. It can't guarantee it will.

#Moving from Google Photos

Immich's feature list doesn't include a Google Takeout importer. The tool people search for alongside it is immich-go, a separate community project (v0.32.0, June 2026). Its upload from-google-photos command reads the Takeout zips and matches photos with their JSON metadata, keeping albums, descriptions and locations. For the wider set of options, see self-hosted Google Photos alternatives.

#Does Immich cost money?

The software is free and AGPL-licensed. FUTO sells optional product keys to support development, and the FAQ asks resellers to send customers to buy them directly.

#Backups

Immich backs up its own database automatically. That covers the database, not your photo files, so the upload folder needs its own copy on another disk or machine. The docs have a full backup and restore page.

#Who it suits

Immich is the right pick if you want your phone's photos off Google or iCloud and onto hardware you own, and you have 8GB or more to give it. If you'd rather index a folder archive you already keep in order, compare it with PhotoPrism in Immich vs PhotoPrism. For what else fits on a Pi, see the best self-hosted apps for a Raspberry Pi.