unmark

Remove pen/pencil markings from a PDF - fully offline. No AI, no network. Your files never leave your machine.

pip install unmark
unmark erase "marked.pdf"            # -> "marked clean.pdf"

erase - the main feature

Strips the markings and hands back a clean PDF. It picks the right path automatically:

• Digital annotations (marks added in GoodNotes / Preview / Adobe, etc.) → deleted losslessly; the original page is restored exactly, text stays text.
• Marks baked into a scan → colour ink (red/blue/green pen) is removed by channel-dominance detection that leaves the printed text untouched. Options: --pencil (light-grey graphite), --lines (experimental underline erase).

unmark erase "scan.pdf" out.pdf --pencil     # also remove light pencil
unmark erase "scan.pdf" --sat 10             # remove more colour (may leave less)
unmark erase "notes.pdf" --lossless          # PNG pages instead of JPEG

What erase can and can't do (honest)

reflow - experimental

⚠️ Experimental. reflow OCRs the printed text (offline, via RapidOCR) and re-typesets a fresh PDF. On clean, single-column papers it's decent. On dense, two-column, or heavily marked papers it is unreliable - it misreads circled option markers, over-splits questions, and leftover dark handwriting bleeds into the text. Always proofread. Offline OCR simply can't match a vision model on messy scans; that's a hard ceiling, not a bug.

pip install "unmark[ocr]"                     # adds the offline OCR engine
unmark reflow "scan.pdf" --columns 1          # single-column works best
unmark reflow "scan.pdf" --pages 1-4          # limit to some pages

How it works

erase: PyMuPDF renders each page → numpy colour/pencil masking → repackaged as a compact greyscale PDF (or the annotation layer is simply deleted). reflow: colour-clean → split columns → RapidOCR → parse → re-typeset (Story API).

License

MIT.