Suggest Lychee photo albums with separate vision and synthesis models
At a glance
A developer builds a local organizer that describes photos and suggests three albums for human selection. This workflow is listed under Local AI, Computer Vision, Photo Management.
- Agents / tools
- Claude Code, Lychee AI Organizer, Lychee, Ollama, Qwen
SourceBuilding a tool for these workflows? See weekly rates.
What they did
1) Have a Lychee MySQL library and database write access. Add nullable `_ai_description` (TEXT) and `_ai_description_ts` (TIMESTAMP) columns to `photos` and `base_albums`. 2) In Ollama, pull `qwen2.5vl:3b` and `qwen3:8b`. On macOS, run `brew install cdzombak/oss/lychee-ai-organizer`. Download `config.example.json` from the `cdzombak/lychee-ai-organizer` GitHub repository as `config.json`; fill in the database credentials, Ollama endpoint, image and synthesis model names, and Lychee base URL. 3) Run `lychee-ai-organizer -config config.json` and open `http://localhost:8080`. Click Describe Photos, then Describe All Albums. 4) The vision model describes photos; the synthesis model summarizes albums and selects three candidate albums for each unsorted photo. 5) Browse the filmstrip and click the appropriate suggested album to move the photo, then continue to the next image.
How it went
The author built the application with Claude Code in a few hours, making only minor manual code edits.
Worth knowing
It writes directly to the Lychee database. Large albums required batched summaries after the synthesis model's context window truncated inputs.