Local Ollama vision models generate Shutterstock-ready metadata (description, keywords, categories, editorial flag, mature content, illustration) and append it to CSV.

• Python 3.10+ and Ollama running with a vision model (tested on llama3.2-vision).
• Install deps (prefer a venv): python -m pip install -r requirements.txt

Directory run (recommended defaults):

python3 image_analyzer.py \
 --dir path/to/images \
 --csv shutterstock.csv \
 --base-url http://localhost:11434/ \
 --model llama3.2-vision \
 --num-predict 800 --top-k 150 --max-retries 3 \
 --resize-max 1024 --resize-quality 85

Single image:

python3 image_analyzer.py --image path/to/img.jpg --csv out.csv --base-url http://localhost:11434/

Options:

• --recursive to include subfolders.
• --prompt-file to override the default prompt.
• --hint to add context for all images.
• --no-fallback disables the safer second pass on failures; --max-retries sets initial attempts (fallback uses cooler options).
• --no-progress for quiet output.

• Caption: one concise sentence <200 chars, no filler/ellipsis/repeated adjectives; prefer exact names when clear.
• Keywords: 7–50 unique, no placeholders/HTML/"...", max 30 after dedup.
• Categories: 1–2 from the fixed list only.
• Booleans: editorial, mature_content, illustration.

• Keyword post-processing (dedup, drop junk, limit length) and category validation.
• Optional resizing/caching (.cache/resized) to improve stability/speed on large images.
• Fallback pass with safer generation options for initial failures.
• Prompt compliance checker for generated CSVs.

python3 -m unittest tests/test_image_analyzer.py

• Generated artifacts (shutterstock.csv, .cache/resized/, sample image folders) are git-ignored by default.
• Use --no-fallback or different options if you need to mirror upstream behavior exactly.