An offline brain for your Nextcloud media server

What it does

Recognize is a Nextcloud app that scans your media library and applies tags without sending data anywhere. It groups faces, identifies objects and landmarks, classifies music genres, and labels human actions in video. Everything runs locally on the server via bundled TensorFlow.js models, and it feeds results into Nextcloud’s built-in Collaborative Tags so apps like Memories and Audio Player can browse the auto-generated categories.

The interesting bit

The project bundles several specialized pre-trained models—EfficientNet for objects, face-api.js for faces, Musicnn for audio, and MoViNet for video—into a single package that drops into a standard Nextcloud instance. The maintainers also publish an “Ethical AI Rating” for each model, noting whether training data is freely available; most get a green check, though the music genre classifier gets a yellow flag for opaque training data.

Key highlights

• Fully on-premise inference using bundled TensorFlow.js—no cloud API calls.
• Recognizes faces, objects, landmarks, music genres, and video actions.
• Integrates with Nextcloud’s built-in Collaborative Tags and apps like Memories.
• Publishes per-model “Ethical AI Ratings” based on open-source training and inference stacks.
• Supports both native (x86_64/AVX) and WASM fallback modes for arm64 or musl systems.

Caveats

• The app is in “limited maintenance” mode: critical bugs are fixed, but new features rely on community pull requests.
• It cannot process end-to-end encrypted files, since the server cannot read them by design.
• The object recognition model alone weighs 1 GB, and the app requires roughly 4 GB of free RAM to operate.

Verdict

Worth a look for Nextcloud users who want private, automated media organization without third-party APIs. Give it a pass if you need active feature development or if your server is tight on RAM.