What Is Immich?
Immich is a self-hosted application for backing up, browsing and searching photos and videos from your phone on a server you control. The GitHub README calls it a "high performance self-hosted photo and video management solution"; the code is open source under the AGPL-3.0 licence.
The App Store lists FUTO Holdings, Inc. as the seller. The apps back up a phone's camera roll to your server, and the web interface adds albums, sharing, a map and administration. The docs describe the AI parts as a separate machine-learning service the server sends images to.
Get It
Immich is a server you install yourself; the documented route is Docker Compose, and the iPhone and Android apps are ordinary store installs that connect to that server.
Option | How to get it |
|---|---|
| Server (Docker Compose) | Install guide: download two files, edit one, run one command |
| iPhone and iPad | App Store (free, iOS 15.0 or later) |
| Android | Google Play, or the APK in the GitHub releases |
| Browser | Any browser at your server's address on port 2283, per the docs |
| Website | immich.app |
| Privacy policy | immich.app/privacy-policy (last updated 6 April 2026) |
| Source code | github.com/immich-app/immich |
| Demo | demo.immich.app, with the login in the README |
Version reviewed: v3.3.5, published on 21 October 2026 per the GitHub release and the App Store record; the Google Play page shows an update on 20 October 2026. Installing needs a terminal and Docker, so this is a developer-style setup rather than a double-click install. The docs ask you to keep the mobile app and the server on the same major and minor version.
Requirements
The docs ask for at least 6 GB of RAM and 2 CPU cores, and recommend 8 GB and 4 cores.
Memory
- What the sources say:
- Minimum 6 GB, recommended 8 GB; with only 4 GB you can run it with machine learning disabled
CPU
- What the sources say:
- 2 cores minimum, 4 recommended; amd64 and arm64; the ML container needs the x86-64-v2 level since v3
Operating system
- What the sources say:
- 64-bit Linux recommended; macOS and Windows via Docker Desktop are possible but discouraged, with limited support
Storage
- What the sources say:
- Unix-compatible filesystem; thumbnails and transcodes add 10-20%; the database should sit on a local SSD, not a network share
Docker
- What the sources say:
- Docker with the Compose plugin; the old docker-compose command is not supported
GPU
- What the sources say:
- Optional for the ML container; the docs call the acceleration experimental
Phone
- What the sources say:
- iOS 15.0 or later per the App Store; Android 8.0 (API level 26) or later per the app build file
These minimums come from the project's docs; PromptQuorum did not measure indexing time or memory use on any hardware.
How to Get Started
These steps follow the project's Docker Compose guide; PromptQuorum has not run them.
- 1Download the compose file
Why it matters: Create a folder, then download docker-compose.yml and the example.env template from the latest release into it. - 2Edit the environment file
Why it matters: Set UPLOAD_LOCATION to a folder with enough space and change DB_PASSWORD to a new value using only letters and digits. - 3Start the containers
Why it matters: Run docker compose up -d in that folder, open the server in a browser and register; the first user becomes the admin. - 4Connect your phone
Why it matters: Install the app, enter your server address, pick the albums to back up and watch progress in the Job Queues tab.
Key Features
Every item below comes from the project's README or documentation; none has been tested by PromptQuorum.
- Backup. Automatic and background backup from the phone, selective albums, duplicate prevention and a command-line uploader.
- Library. Albums, shared albums, partner sharing, public links, a map, memories, stacks, tags, a folder view and RAW support.
- Search. Contextual (CLIP) search plus filters for people, places, file name, description, tags, camera, dates and star rating; the docs list text in images (OCR) as a search type.
- Accounts. Multiple users, OAuth sign-in and API keys; the first registered user is the admin.
- Import. immich-go for Google Takeout and External Libraries for archives that already sit on disk.
AI Features and Where They Run
The AI is image-based: CLIP models for smart search and models for face detection and recognition, all run by a machine-learning container; no language model is involved.
- Smart search. Freeform queries are matched against CLIP embeddings stored in Postgres through the VectorChord extension; you pick the model under Administration, and the default is the fastest one.
- Languages. The docs say nllb models expect queries in the language set in user settings, while xlm and siglip2 models handle text in any language.
- Faces. Detection runs on the preview image, a recognition model produces embeddings, and DBSCAN clustering groups them into people you can name.
- Hosting. The built-in ML service is configured by default, and the FAQ says it can run on a more powerful machine instead.
- GPU. CUDA, ROCm, OpenVINO, ARM NN and RKNN backends exist; the docs call the feature experimental, CUDA the most reliable and ROCm new.
Data Flow and Privacy
The docs say that once a smart-search model is downloaded, Immich needs no network connection unless the version check is enabled, and that reverse geocoding runs on your own server from GeoNames data. Photos and the database therefore stay on your hardware unless you share them.
Outside requests named in the sources: model downloads from Hugging Face, the optional version check, which the server source sends to Immich's own version service rather than to GitHub, and map tiles. The privacy policy, last updated on 6 April 2026, lists a Tile Server and a Version Service that log request metadata (the requested tile or the user agent, the IP address and timing); it says the tile server is optional and can be disabled. Server-side email goes through an SMTP server you configure. The Google Play listing declares that no data is shared with third parties, that photos and videos and device or other IDs may be collected, and that data is encrypted in transit.
📌Note: These are the project's declarations, not audit results. The docs name GitHub for the version check, while the server source and the policy name Immich's own Version Service, and PromptQuorum has not inspected network traffic or read the rest of the source code.
Who Should Use It
- Self-hosters with a Docker server or NAS who want a private replacement for a cloud photo service.
- Families and small groups who want one library with separate accounts, shared albums and phone backup.
- People with large camera-roll archives who need search by content and by person.
What We Could Not Verify
- AI quality. Search relevance, face-grouping accuracy and speed were not tested; the docs' benchmark figures come from one test system.
- Backups. The docs warn that the database holds only metadata and that you must back up the upload folder yourself, and the README asks for a 3-2-1 backup plan.
- Trash behaviour. The FAQ says original files that were trashed are deleted when the trash is emptied in Immich.
- Purchase terms. The purchase page could not be read, so prices are not stated; the docs say purchasing is voluntary support.
- Store declarations. The Google Play data-safety summary is the developer's own declaration, and the App Store privacy label was not read.
Competitors and Alternatives
The directory has two other photo-library tools with a review, one self-hosted and one cloud-based.
PhotoPrism
- Platforms:
- Server (Docker), browser
- Price / license:
- Free Community / AGPL-3.0
- Key difference:
- Self-hosted like Immich, but web-app only; its phone apps come from other developers
Articles about PhotoPrism (1)
- PhotoPrism Review: Self-Hosted AI Photo Library for Your ServerUpdated October 24, 2026
Ente Photos
- Platforms:
- iOS, Android, desktop, web
- Price / license:
- Free 10 GB, paid plans / AGPL-3.0
- Key difference:
- Encrypted cloud storage with on-device AI; no server of your own
Articles about Ente Photos (1)
- Ente Photos Review: Encrypted Backup with On-Device AI SearchUpdated October 24, 2026
Confirm each app's current price, license and platforms on its own page.
Frequently Asked Questions
Is Immich a local LLM?
No. The docs describe CLIP image search and face recognition, not a language model, and the sources read contain no chat or caption-writing feature.
Does Immich work without a GPU?
The docs list GPU acceleration as optional and experimental, and the minimum requirements name only RAM and CPU cores; speed on a CPU was not measured.
Can I move over from Google Photos?
The introduction page points to immich-go for Google Takeout exports; PromptQuorum did not try a migration.
How do I update Immich?
The install guide refers to separate upgrade instructions, and the FAQ asks you to keep the app and the server on the same major and minor version.
Where are my files kept?
In the folder set as UPLOAD_LOCATION; the database in the Postgres container holds metadata and user information only.
Verdict
Immich is a free, actively updated way to keep a searchable photo library with phone backup on hardware you control, with content search and face grouping that run on your own server. Against that, setup needs Docker and an environment file, GPU acceleration is labelled experimental, and none of the AI quality claims could be checked here. It suits people who already run services at home or on a NAS; anyone who wants a double-click app or a language-model assistant should look elsewhere.
Sources
- Immich website — platforms, links and privacy policy, checked 26 October 2026.
- Immich documentation — requirements, install guide, smart search, facial recognition, system settings and FAQ, checked 26 October 2026.
- Immich privacy policy — locally stored data, tile server and version service, last updated 6 April 2026.
- Immich on GitHub — README, releases and commit activity, checked 26 October 2026.
- Immich on the App Store and Google Play — seller, version and update dates, checked 26 October 2026.