Danbooru Tag Search

Standalone Danbooru-style tag search using WD Tagger v3. All in memory — no cache, no database.

Setup

.\setup.ps1
.\.venv\Scripts\Activate.ps1

Usage

# Copy matches to destination
python danbooru_search.py search /path/to/images \
    --tag solo_focus --threshold 0.5 --dest /path/to/output

# Multiple tags (AND — all must match)
python danbooru_search.py search /path/to/images \
    --tag 1girl abs --threshold 0.5 --dest /path/to/output

# Multiple tags (OR — any match)
python danbooru_search.py search /path/to/images \
    --tag 1girl solo_focus --any --threshold 0.5 --dest /path/to/output

# Move instead of copy
python danbooru_search.py search /path/to/images \
    --tag solo_focus --threshold 0.5 --dest /path/to/output --move

# Dry run (preview)
python danbooru_search.py search /path/to/images \
    --tag solo_focus --threshold 0.5 --dest /path/to/output --dry-run

Options

Flag	Description	Default
--tag, -t	Tag(s) to search for (repeatable)	required
--threshold	Minimum tag confidence score	0.5
--dest, -d	Output directory	required
--any	OR logic (any tag matches), default is AND (all must match)	False
--move	Move instead of copy	False
--dry-run	Preview without copying/moving	False

Notes

• Uses p1atdev/wd-swinv2-tagger-v3-hf (WD Tagger v3, SwinV2-based, trained on Danbooru)
• All tagging is done in-memory — no SQLite cache, no disk persistence
• Model is lazy-loaded on first use