LibrePhotos runs its face, caption and search models on ONNX Runtime, and asks for 8 GB of RAM to do it comfortably
A self-hosted open source photo management service.
At a glance
- What is it?
- LibrePhotos is an MIT licensed, self-hosted photo manager with a Django 6 backend, a React 19 web client and an Expo mobile app, now merged into one monorepo. Its machine learning pipeline drops PyTorch for ONNX Runtime, but the memory floor, the fast release cadence and the mobile build's pinned dependencies are what an operator has to plan around.
- Who is it for?
- LibrePhotos fits people who want a self-hosted photo library with face clustering, generated captions, semantic search, multiple users and OpenID Connect sign-in, and who can give it a machine with 8 GB of RAM or more. On a 4 GB box it starts, but the machine learning features compete for memory.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 2 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 13, 2026, and from our analysis. They are not legal advice.
Star history
- LibrePhotos/librephotos8,158 ★ · +85
| Repository | 2026-09-13 | 2026-11-13 |
|---|---|---|
| LibrePhotos/librephotos | 8.1k | 8.2k |
85 Stars in the last 61 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
Five former repositories now live under apps/ and deploy/
LibrePhotos used to be split across five repositories, and they have been folded into one monorepo on the dev branch. apps/backend/ holds the Django 6 API, the machine learning pipelines and the background jobs, and was the old librephotos repository. apps/frontend/ is the React 19 and Vite web client with i18next localization, formerly librephotos-frontend. apps/mobile/ is the Expo app for Android and iOS, built offline-first, formerly librephotos-mobile. apps/docs/ is the Docusaurus site published to docs.librephotos.com, formerly librephotos.docs. deploy/ carries the Dockerfiles, Compose configs, proxy and Kubernetes manifests that used to sit in librephotos-docker. Commit history from all five was preserved, so git log --follow on a file under apps/ still reaches back across the move. For anyone who bookmarked issues or forks of the old repositories, the practical effect is that code now has to be found by path in one tree, and pull requests for the web client, the docs and the deployment files all land in the same place.
Every model runs on ONNX Runtime, with no PyTorch
The machine learning stack is the part that sets LibrePhotos apart from a plain gallery, and it is built so that no PyTorch install is needed: ONNX Runtime executes every model. Face detection uses InsightFace. Face classification and clustering use hdbscan clustering and a scikit-learn multilayer perceptron, both ported to numpy, which is how the app can ask you to name a person once and then suggest that name for the rest of the cluster. Captions come from LFM2.5-VL in its 450M ONNX build, and the captioning prompt is fed the recognised people and the place, so captions can mention who and where rather than only what. Tags, semantic search and similar photo lookup use OpenCLIP ViT-B/32 trained on DataComp-XL, which carries an MIT licence. A CLIP ViT-B/32 model with a FAISS index also appears under semantic search, so two CLIP models are in play. Reverse geocoding for places and the map goes through geopy. The result is a search box that accepts phrases such as dog on a beach next to metadata search.
The 4 GB minimum leaves the ML features short of memory
The published minimum is 4 GB of RAM, 2 CPU cores and 10 GB of storage on top of the photo library itself, on any Docker compatible OS or on Windows. The recommended machine is 8 GB or more of RAM, 4 or more cores, an SSD, and Linux. The project is direct about why the two columns differ: face recognition, scene classification and image captioning are memory intensive, and 8 GB or more is strongly recommended for smooth operation. The consequence for a small home server is that meeting the minimum does not mean the ML side runs smoothly; the minimum column is the floor for the service, and the features that make LibrePhotos worth installing over a plain file gallery are the ones the project ties to the recommended column. If your only hardware sits at the minimum, expect to trade comfort for capacity, or put LibrePhotos on a machine that meets the recommended figures and keep the small box for something else. The storage line also matters: 10 GB is for the application, its models and its data, and your photos come on top.
Five deployment routes, each with a different footprint
There is no single install command; the project offers five routes and links each to a step by step guide on docs.librephotos.com. The single container route runs one image that serves both the API and the web app, and it is the one recommended for most people. The Docker Compose route splits backend, frontend and proxy into separate containers and is the route that comes with optional NVIDIA GPU acceleration. Kubernetes users get Kustomize manifests in deploy/k8s. Unraid has its own guide. The Windows standalone route is one folder containing librephotos.exe that needs neither Docker nor Python, and a release tag named windows-wheels-2026.10.22 was published on 2026-10-22. Nginx is the reverse proxy in the stack. Once the app is up, interactive API documentation is served at http://localhost:3000/api/swagger and http://localhost:3000/api/redoc. Choosing a route is a real decision rather than a formality: moving from the single container to Compose to gain GPU acceleration means a different container layout to maintain.
RAW+JPEG pairs collapse into one photo through LibRaw and ExifTool
On the import side LibrePhotos scans existing folders and accepts uploads from the web and mobile apps. It reads photos in common formats including RAW, HEIC and JPEG XL, plus videos and live photos. RAW conversion goes through LibRaw via rawpy, video conversion through FFmpeg, and Exif handling through ExifTool. Files that a camera writes as a RAW+JPEG pair are kept as one photo instead of two, burst and bracket sequences are stacked, and duplicates are detected. The timeline groups photos by day and place and adds favorites, recent additions and memories, and the app generates event albums with names like Saturday in Lisbon. Data lives in PostgreSQL, with Django-Q2 as the task queue for the backend's background jobs. Because the scanner reads folders you already have, a library you organise by hand can stay where it is rather than being moved through the upload path, and the web and mobile uploads become one more source next to it.
Multiple users, shared albums and OpenID Connect sign-in
LibrePhotos is built for more than one person. It supports multiple users, shared albums and public links, and it can delegate login to an identity provider through OpenID Connect single sign-on. For a household or a small team that already runs an identity provider, OIDC means one account per person instead of another password store. Public links are the way to share an album with someone who has no account. The web client is React 19 with TypeScript, using Mantine for components, TanStack Router and TanStack Query, MapLibre GL for the map view and i18next for localization, with translations collected through Weblate. The Expo mobile app for Android and iOS is offline-first and uploads into the same server. The API that both clients use is Django REST Framework, documented through the Swagger and ReDoc pages, which also gives scripted access to the library for anyone who wants to automate imports or exports from their own tools.
The mobile build depends on hand-pinned Expo and React versions
The root package.json is a private npm workspace named librephotos-monorepo, and only packages/* and apps/mobile are workspaces; apps/frontend and apps/docs keep their own independent Yarn installs and are deliberately excluded. Contributors therefore work with npm for the mobile app and Yarn for the web client and docs. The root file also pins dependencies by hand, with comments explaining why. expo-router is declared at the root only to force npm to hoist it; without that, the root Expo CLI's typed route generator cannot resolve expo-router/_ctx-shared and expo start crashes outright. expo is pinned to the same range as apps/mobile because expo-router peer-depends on expo at any version, and npm would otherwise put the latest Expo at the root. react and react-dom are pinned to 19.2.3, the React version that Expo SDK 57 and react-native 0.86 ship with, because jest-expo@57 pins [email protected] and jest-expo hard-fails on any mismatch. The cost is that an Expo SDK upgrade means updating several pins in lockstep, and missing one breaks the mobile toolchain rather than producing a warning.
Releases arrive days apart, so upgrades need reading first
LibrePhotos moves quickly. Release 1.5.0, titled October 2026, was published on 2026-10-28, and 1.5.1 followed on 2026-10-31 under the title Faces in the right place, upgrade step fixed on SQLite. So 1.5.1 is a fix release, three days after a feature release, for face placement and for the upgrade step on SQLite. The stack lists PostgreSQL as the database, so the SQLite mention means at least one deployment path runs on SQLite, and the release title does not say which. If you upgraded to 1.5.0 on SQLite, the 1.5.1 notes are the place to check before doing anything else. Testers are asked to run the dev tag and update it regularly, which is the opposite of what a household photo archive should do. For a library you care about, pin a numbered release, back up the database and the data volume before each upgrade, and skip a fresh .0 release until its follow-up appears. The last push to the dev branch was on 2026-11-03.
Editorial conclusion
LibrePhotos fits people who want a self-hosted photo library with face clustering, generated captions, semantic search, multiple users and OpenID Connect sign-in, and who can give it a machine with 8 GB of RAM or more. On a 4 GB box it starts, but the machine learning features compete for memory. Before installing, pick one of the five deployment routes deliberately, back up the database, and read the release notes of the version you pull, since 1.5.0 and 1.5.1 shipped three days apart with an upgrade fix for SQLite.
Frequently asked questions
How does LibrePhotos compare with PhotoPrism and Immich?
The LibrePhotos README does not include a comparison with PhotoPrism or Immich. On its own terms, LibrePhotos runs every model on ONNX Runtime without PyTorch, offers a Windows standalone build with no Docker or Python, and supports OpenID Connect single sign-on.
Is LibrePhotos a free alternative to Google Photos?
LibrePhotos is free, MIT licensed and self-hosted, with a timeline, face clustering, generated captions, semantic search and shared albums. You run it on your own hardware, which needs at least 4 GB of RAM, with 8 GB or more recommended.
Can LibrePhotos store photos from a phone?
Yes. LibrePhotos has an offline-first Expo mobile app for Android and iOS that uploads to your server, and it also accepts uploads from the web app and scans existing folders.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/librephotos-librephotos)Star history chart for your README
[](https://hysenlabs.com/projects/librephotos-librephotos)Was this useful?