TingdeLiu PDF Figure Extractor
Extract figures and tables from PDF files using TF-ID (Florence2-based) object detection.
Environment
- Conda env:
TF-ID - Dependencies:
pdf2image,transformers,pillow
Usage
Windows (recommended — avoids encoding issues)
Invoke the bundled .bat from this skill's scripts/ directory:
cmd /c "scripts\extract_images.bat" "<pdf_path>" "<pages>"
Examples:
# Extract all pages
cmd /c "scripts\extract_images.bat" "paper.pdf" ""
# Extract specific pages
cmd /c "scripts\extract_images.bat" "paper.pdf" "2,5,10,11"
# Extract page range
cmd /c "scripts\extract_images.bat" "paper.pdf" "3-7,10"
Direct Python (cross-platform)
conda run -n TF-ID python scripts/extract_images.py <pdf_path> -o <output_dir> -t figure table --pages <pages>
Paths above are relative to this skill's directory. When the skill is installed under
~/.claude/skills/pdf-figure-extractor/, run the commands from that directory, or prepend the full skill path.
Parameters
| Flag | Description | Default |
|---|---|---|
pdf_path |
Input PDF file | required |
-o / --output |
Output directory | ./images |
-t / --type |
figure, table, or both |
figure |
--pages / -p |
Pages to process, e.g. 1,3,5 or 2-6,10 |
all pages |
--model |
TF-ID model ID | yifeihu/TF-ID-base |
Output
<output_dir>/figure_01.png,figure_02.png, ... — cropped figures<output_dir>/table_01.png,table_02.png, ... — cropped tables (if requested)<output_dir>/image_index.json— machine-readable index<output_dir>/image_index.md— human-readable index with page numbers
Workflow
- Ask user for PDF path (and optionally: output dir, types to extract, page range)
- Run extraction via bat script (Windows) or
conda run(other OS) - Read
image_index.mdto report what was extracted - Show user the output directory and extracted items summary
Notes
- The bat script hardcodes output to
.\imagesrelative to the skill directory; for custom output dir use direct Python invocation - When page list is unknown, extract all pages first, then let user filter
- TF-ID detects layout elements — it works best on typical academic paper layouts
- For large PDFs, specify target pages to save time and memory