A few years ago, I built a Local Image Search app for managing my painting motives and reference images.
My new Image Tag Manager is the replacement, built with Tauri, React, and Rust.
Thanks to using Qwen3-VL-2B-Instruct, the results are much better than with whatever i was using back in the day. The model runs locally on the CPU and produces five to eight tags and a caption per image.
The results go into an SQLite DB, so searching is fast.
The app indexes multiple folders and resumes pending work after a restart. Search covers tags, captions, filenames, and categories. Words combine with AND, while #tags match exact tags. I can inspect images in a large viewer and rename or merge tags across the library.
Unlike the old tool, this version stores tags in its database rather than image metadata. I need to back up the database along with the pictures.
I’m still indexing my library. Next I’ll run both models on the same sample and check which obvious subjects they miss, how often they invent tags, and how long classification takes. More tags aren’t useful if they’re wrong.