Show HN: Image Categorizer – sort a photo library with local vision models

Sorts a folder of photos into categories using local AI models served by Ollama. A vision model describes every photo and flags accidental shots; a text model sorts the photos into categories you keep and refine across runs. You review the result in a browser report and export a script that moves the files into folders, writes each category into the photo's caption, or both. No photo leaves your network.

Built for camera rolls full of accidental shots, near-duplicates and screenshots, where reviewing every photo by hand is not practical.

1. Describe. A vision model looks at each photo, writes a short description, and decides whether it is an accidental or failed shot (motion blur, pocket shot, no subject). With--describe-only (see below), descriptions are saved as they go and an interrupted run resumes where it stopped.
2. Sort. A text model assigns every photo to a category using the descriptions only, so re-sorting takes minutes and never looks at the images again. Each category has a rule describing what belongs in it, and the list is saved and reused for every library.
3. Review. A report in the browser lets you move photos between categories, trash them, and rename, merge or create categories.
4. Apply. The report exports a bash script: move each photo into a folder named after its category, writeCategory: <name> into its caption (searchable in the Photos app on Mac and iPhone), or both. Nothing is deleted.

The tool itself never moves, renames or deletes a photo.

With an AI coding agent. Open the repository in Claude Code (or a similar agent) and ask it to follow SETUP.md. It checks the machine, installs what is missing, configures Ollama and the models, and runs a smoke test on the images in samples/.

By hand. You need uv and an Ollama server, on this machine or another one on your network.

git clone https://github.com/DrBenedictPorkins/image-categorizer.git
cd image-categorizer
uv sync

ollama pull qwen3.6:35b            # vision model (describe)
ollama pull mistral-small3.2:24b   # text model (sort)
cp .env.example .env               # set OLLAMA_HOST if Ollama runs elsewhere

uv run python main.py samples --provider ollama

The last command sorts the eleven sample images and opens the report. Smaller model pairs for 8-16 GB GPUs are listed in SETUP.md.

Describe once, then sort as often as you like:

uv run python main.py "/path/to/photos" --description-provider ollama --describe-only

uv run python main.py "/path/to/photos" --categorization-provider ollama \
    --categorize-from "/path/to/photos/descriptions_only.json" --plan-categories --max-categories 12

uv run python main.py "/path/to/photos" --categorization-provider ollama \
    --categorize-from "/path/to/photos/descriptions_only.json" --max-categories 12

For reference: 1,195 photos took about 80 minutes to describe and 8 minutes to sort with qwen3.6:35b and mistral-small3.2:24b on an RTX 4090.

• Photos grouped by category; the rail on the left shows counts and accepts drops
• Move a photo by dragging it or with its category menu, which also offers the model's suggested categories
• Rename, merge (rename to an existing name), create and remove categories; edit each category's rule
• Trash with restore; move a whole category to Trash in one step
• Full-size preview with the model's description; arrow keys browse the category
• Save categories downloads the category list, with your edits, for future runs
• Export script writes the organize script

Writing captions requires exiftool. An existing caption is kept; a previous Category: part is replaced, so the script can be run again.

Categories are saved in ~/.config/image-categorizer/categories.yaml (override with CATEGORIES_FILE) and reused for every library:

categories:
  - name: Portraits
    rule: Photos of people looking at the camera, including selfies. Not group photos.
    status: kept

The first run builds the list from the photos. Later runs sort into the saved list; only when enough photos fit none of the rules does the model propose a new category, marked new in the report until you save the list. --max-categories caps the list: once it is full, photos that fit nothing go to Unsorted.

When a category holds photos that belong elsewhere, click the re-sort button on that category and describe what went wrong.

uv run python main.py --resort ~/Downloads/resort.json

The text model turns the note into rule changes and asks you to approve them, shows how 10 sample photos would move, then re-sorts only that category. Photos that fit no other category stay put. Moved photos are marked Re-sorted in the rebuilt report, and edits made in the report before exporting are kept.

Settings live in .env; .env.example lists them.

All command-line options: uv run python main.py --help and SETUP.md. Supported formats: .jpg, .jpeg, .png, .gif, .bmp, .webp, .heic, .heif; videos are skipped. The report shows HEIC photos through JPEG previews kept in a hidden .image-categorizer-previews folder.

--provider huggingface runs BLIP-2, LLaVA or Flan-T5 in-process without Ollama (docs/HUGGINGFACE_PROVIDER.md), and --categorization-provider keyword sorts by keyword matching. Both predate the saved categories and re-sort features, and are less maintained.

MIT. See LICENSE.