AI-powered metadata tagger for your photo and video library.
Meta-Analyzer is a local-first desktop app that looks at your images and videos with a vision LLM and writes descriptive tags (and, for videos, a title, description and genres) straight into the file's metadata — so your media becomes searchable in tools like Plex, digiKam or your OS file search.
It runs entirely against a local, OpenAI-compatible model server (e.g. LM Studio) by default — your files never leave your machine. An optional API key lets you point it at a private/cloud server instead.
Built with Tauri 2 · Svelte 5 · Rust.
- Photos & videos — two pipelines sharing one engine.
- Photos: tags written to EXIF / IPTC / XMP.
- Videos: title, description, genres and tags written to MP4/MKV/AVI
containers via
ffmpeg; frames sampled from an embedded cover or evenly across the clip.
- Local-first — talks to any OpenAI-compatible
/v1/chat/completionsendpoint (LM Studio, etc.). Optional Bearer API key for private/cloud servers. - Content-type profiles — General and Custom are built in; every other
category is an editable JSON rule file (
image_*/video_*) that is discovered at runtime. Drop in your own or share them — a Rules folder button opens the directory. - Prompt Builder — per-profile control over tag count, output language (with a strict translation directive + English fallback), vocabulary mode (Strict / Recommended / Optional), extra custom vocabulary and free-text custom instructions.
- Batch processing — worker pool with configurable concurrency, pause / resume / stop, live token-streaming "thinking" panel per worker, and an ETA timer.
- Review or automate — auto-apply results, or edit the title, description, tags and genres inline before writing.
- Quality-of-life — light/dark theme, drag & drop, inline thumbnails, model favourites, toasts, and reveal-in-Explorer.
- A running OpenAI-compatible vision model (default: LM Studio with a vision-capable model loaded).
ffmpeg/ffprobeon yourPATH(or next to the executable) — required for video frame extraction and metadata writing. Not needed for photo-only use.- Windows (primary target). The stack is cross-platform via Tauri; other platforms are untested.
Output quality depends heavily on the model you run. Notes from testing:
- Vision is mandatory — the model must be able to see images (a vision/multimodal model). Text-only models will not work.
- Reasoning is recommended — reasoning-capable models produce noticeably more reliable results and stick to the required JSON format far better.
- You own the model configuration — Meta-Analyzer only sends the requests. The number of parallel workers is ultimately bounded by your LM Studio (server) settings, and it is up to you to load and configure your model so it runs stably under batch load (context length, GPU layers, parallel request limits, etc.).
Recommended — reliable JSON output + strong analysis:
- Qwen3.5 9B — Q6 or higher
- Mistral 3 14B — Q4 or higher
Faster, but notably less accurate:
- Qwen2 7B (Q8), no reasoning — fast, but makes more mistakes and does not always return valid JSON.
The prompt generator and the built-in presets were tuned and tested mainly against Qwen3.5 and Mistral 3. Cloud providers are untested for output quality.
npm install
npm run tauri devPoint the app at your model server URL (default http://localhost:1234/v1),
pick a model, choose a content type, and start.
npm run tauri build| Layer | Tech |
|---|---|
| Shell | Tauri 2 (Rust) |
| Frontend | Svelte 5 (runes) · TypeScript · Vite · TailwindCSS |
| Backend | Rust — reqwest (SSE streaming), tokio, little_exif / img-parts / image, ffmpeg (external) |
| Model API | OpenAI-compatible /v1/chat/completions |
- User Guide — install, first-run setup, tagging photos & videos.
- Visual walkthroughs — Interface · Tagging Photos · Tagging Videos
- Developer docs — architecture, build, data flow, prompt builder, metadata format.
Meta-Analyzer is free software licensed under the GNU General Public License v3.0 or later.
It is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY. See the LICENSE file for the full terms.