figcrop 0.2.0

High-quality figure extraction from dense research-paper PDFs

figcrop

Extract figures from research-paper PDFs by real figure number.

figcrop uses MinerU's PP-DocLayoutV2 layout model for visual-region detection, then adds its own PDF-text and geometry logic to group panels into whole figures, remove captions, and trim page furniture. It is designed for dense semiconductor papers where simple heuristic tools often miss figures or split panels badly.

Features

• Extract whole Fig.N outputs, including multi-panel figures, by the real figure number found in the PDF text layer.
• Keep figure-internal labels, process-flow text, axes, table borders, and panels, while excluding the caption line.
• Output modes:
  • figure default: whole figure body, no caption.
  • panel: split figures into (a) / (b) subpanels when panel labels are available.
  • caption: whole figure body plus the matched caption text.
• Run as a persistent local HTTP server so the layout model is loaded once.
• Use OpenVINO by default for fast local inference; torch backends remain available.
• Trim modes:
  • mask default: fast trim using an ignore mask for captions, page furniture, neighbor frame lines, and thin anti-aliased edge bleed.
  • whiteband: extra local whitespace snapping around the detector bbox. This is slower but useful as a conservative safety mode for difficult crops.
• Handles slide/poster-style cells in some PDFs by clipping giant page-level fallback detections to the caption's local cell.

Requirements

• Python 3.10-3.13. Windows is the main tested environment.
• uv is recommended; the setup scripts install into the project-local .venv.
• First run downloads model weights from Hugging Face.
• OpenVINO GPU is the recommended default. NPU is not currently useful for this RT-DETR layout model.

Setup

Windows:

powershell -ExecutionPolicy Bypass -File setup.ps1 -Device xpu

Linux/macOS:

bash setup.sh

requirements.txt is a manifest for the base packages. Device-specific torch packages are handled by the setup scripts.

Install

Install the released CLI from PyPI:

pipx install figcrop

This exposes:

figcrop help
figcrop extract paper.pdf out --mode caption
figcrop serve
figcrop-mcp

figcrop is the main human/agent CLI. figcrop-mcp is only needed when an MCP client should call figcrop as a registered tool instead of running shell commands.

To install directly from the current GitHub source instead:

pipx install git+https://github.com/tadalab-keio/figcrop.git

Usage

<py> means .venv\Scripts\python.exe on Windows or .venv/bin/python on Unix.

For AI agents, the shortest path is:

<py> figtools.py extract paper.pdf out auto --mode figure

Then inspect the JPEGs and out/figures.json. Do not treat a successful command as a successful crop until representative images have been viewed.

Start the local server:

<py> figtools.py serve auto

Then request crops:

curl -s -X POST http://127.0.0.1:8077/extract \
  -H "Content-Type: application/json" \
  -d '{"pdf":"paper.pdf","out_dir":"out","figs":[1,2],"mode":"caption"}'

Useful request fields:

• figs: real figure numbers to extract, for example [1,2].
• top: fallback positional extraction, for example 2 for the first two visual regions per page.
• mode: "figure" default, "panel" for (a) / (b) subpanels, or "caption" for whole figure plus matched caption text.
• panels: legacy alias for mode: "panel".
• trim_mode: "mask" default or "whiteband".
• caption_mode: legacy alias; "include" means mode: "caption".

One-shot CLI:

figcrop extract paper.pdf out auto
figcrop extract paper.pdf out auto --figs 1,2
figcrop extract paper.pdf out auto --top 3
figcrop extract paper.pdf out_panels auto --mode panel
figcrop extract paper.pdf out_whiteband auto --trim whiteband
figcrop extract paper.pdf out_with_captions auto --mode caption
figcrop help

Output files are JPEG crops plus a figures.json manifest in out_dir.

AI Connectors

• OpenAPI/REST: start figtools.py serve auto and use /openapi.json.
• MCP: install requirements-mcp.txt and run figcrop_mcp.py.

See CONNECTORS.md for Claude/Codex/MCP examples and connector safety notes.

How It Works

1. Render each PDF page at 150 dpi for layout detection.
2. Run PP-DocLayoutV2 through OpenVINO or torch.
3. Keep visual regions labeled image, chart, or table.
4. Read Fig.N / Table N captions directly from the PDF text layer.
5. Assign each region to the nearest same-column caption below it.
6. Union regions with the same figure number into one whole-figure crop.
7. In panel mode, split whole-figure boxes using (a) / (b) label anchors.
8. In caption mode, extend the crop to the matched caption paragraph.
9. Render the page at 300 dpi and crop with the selected trim mode.

The numbering and whole-figure grouping are local geometry logic, not MinerU's full reading-order pipeline.

Performance

For repeated extraction, use the server. In local tests, server mode avoids reloading the model for each request and is much faster than one-shot CLI runs, especially on one-page PDFs.

For occasional use, the one-shot CLI is still practical: most ordinary papers finish in a few seconds on a local OpenVINO GPU setup, so running the server is a convenience rather than a hard requirement.

The default mask trim mode is optimized for speed and is the recommended mode. whiteband is kept as a higher-conservatism option when local whitespace around the detector bbox matters more than speed.

Limitations

• Very dense pages can still confuse region-to-caption assignment. Use top= or mode=panel as fallbacks.
• Some PDFs contain slide/poster grids or decorative page furniture that can look like a large table. figcrop has a local-cell fallback for common cases, but this class of PDF may still need review.
• Outputs are intended for local research workflow use. Always inspect crops when building datasets or publications.

Agent Notes

See AGENTS.md, CLAUDE.md, and CONNECTORS.md for machine-oriented commands, connector setup, verification checks, and commit attribution conventions. In short: choose exactly one output mode (figure, panel, or caption), make montages for visual QA, and use --force-with-lease rather than plain --force if rewriting a pushed history.

License

Apache-2.0. See LICENSE.

Built on MinerU and OpenVINO. See NOTICE for attribution and dependency notes.