paper_figure_extractor

Use this skill when the user asks to extract figures from scientific papers, including arXiv IDs/URLs, local PDFs, and journal article pages such as Nature-style pages.

Hard rules

• Never fabricate figures.
• Never fabricate captions.
• Never fabricate figure numbers.
• Never fabricate panel labels such as a, b, or c.
• Never generate, redraw, reconstruct, or visually invent images.
• Never infer missing panels without evidence.
• If anything is uncertain, mark it uncertain, save it as a candidate, and do not promote it to final FigXX.
• Every output must have provenance pointing to a source file, LaTeX location, HTML location, or PDF page/object location.

Tool usage

Use the bundled paperfig tool rather than visually inventing figures:

paperfig extract INPUT --out OUTPUT --dpi 400 --strict

Python API:

from paperfig import extract_figures
extract_figures(input, outdir, dpi=400, strict=True)

Reliability hierarchy

Always prefer higher-confidence sources. Lower levels must not override higher levels.

1. High: arXiv source LaTeX, parsed from figure environments, \includegraphics, and \caption.
2. Medium: journal HTML <figure> and <figcaption> elements.
3. Low/medium candidate source: PDF embedded image objects.
4. Low: rendered PDF page crops and caption-based crops. In strict mode these are candidates only.

Output contract

The tool writes:

OUTPUT/
  metadata.json
  figures/
    Fig01.png
    Fig01_panel_a.png
  candidates/
    Fig03_candidate_page_5.png
    Fig04_panel_candidate_01.png
  captions/
    Fig01.txt

Only verified source-backed figures and panels belong in figures/. Uncertain images, ambiguous mappings, rendered PDF crops, and visually inferred panels belong in candidates/.

Panel rules

Verified panel labels require source evidence, such as multiple \includegraphics entries inside one LaTeX figure environment or multiple images inside one HTML figure. Caption-only references can support metadata but must not cause visual crops or labels unless alignment is reliable. Visual/grid inference is candidate-only and must not assign a, b, or c.

Reporting results

When presenting results to the user, distinguish verified figures from candidates. Do not describe candidates as extracted final figures. Refer to metadata.json for provenance and confidence.