Auto-Rotate

Automatically deskew and turn upright every page of a PDF — CLI, Python library, and a cross-platform desktop app.

Each page is rasterized, snapped to its correct cardinal orientation (0/90/180/270°) with Tesseract's orientation-and-script detection, then fine-skew corrected with jdeskew (FFT + radial projection), and reassembled into a new PDF. Optionally, a searchable text layer is added with OCRmyPDF.

PDF page
  → rasterize @ DPI (pypdfium2)
  → cardinal orientation snap   (Tesseract OSD, lossless np.rot90)   [--orient, default on]
  → fine-skew correction ±15°   (jdeskew, white-filled corners)
  → binarize to black-on-white  (Sauvola local threshold)            [--clean, default off]
  → rebuild PDF (img2pdf)
  → searchable text layer       (OCRmyPDF)                           [--ocr, default off]

Install

Desktop app (no Python needed) — download the installer for your OS from the latest release: Windows .msi, macOS .dmg (Apple Silicon, macOS 14+), or Linux .deb/.rpm/Flatpak.

The desktop builds are not code-signed (this is a free open-source project without Apple/Windows signing certificates), so the OS warns on first launch:

• macOS: right-click the app → Open (or System Settings → Privacy & Security → Open Anyway). One time only. To skip the prompt entirely, install via pipx below.
• Windows: SmartScreen → More info → Run anyway.

CLI / library / GUI via pipx or pip — no Gatekeeper/SmartScreen prompt, since nothing is downloaded as an app bundle:

pipx install auto-rotate            # the `auto-rotate` command
pipx install "auto-rotate[ocr]"     # + the `--ocr` capability
pipx install "auto-rotate[gui]"     # + the `auto-rotate-gui` desktop app
pip install auto-rotate             # as a library

From source with uv:

uv sync                  # core
uv sync --all-extras     # core + gui + ocr + dev tools

System binaries

On macOS: brew install tesseract ocrmypdf.

Usage

auto-rotate input.pdf output.pdf                  # cardinal orientation + deskew (image-only)
auto-rotate input.pdf output.pdf --no-orient      # deskew only, no Tesseract needed
auto-rotate input.pdf output.pdf --clean          # + pure black-on-white (great for scans)
auto-rotate input.pdf output.pdf --ocr            # + searchable text layer
auto-rotate input.pdf output.pdf --dpi 400 -v     # higher render DPI, log per-page corrections

positional: input, output
--dpi N            rasterization DPI (default 300)
--orient/--no-orient   cardinal-orientation detection (default: on)
--angle-max DEG    max fine-skew angle searched (default 15.0)
--clean            binarize to pure black-on-white (flattens scan background)
--ocr              add a searchable text layer (needs ocrmypdf)
-v, --verbose      log the correction applied to each page

Desktop app

auto-rotate-gui     # if installed via pip/pipx with the [gui] extra

Add one or more PDFs, toggle orient / clean / OCR, pick a DPI, and Run; each file is written next to its source as <name> - upright.pdf, with per-page progress in the log. Orientation and OCR controls are enabled only when Tesseract / OCRmyPDF are detected.

Library

from pathlib import Path
from auto_rotate import deskew_pdf

results = deskew_pdf(Path("scan.pdf"), Path("upright.pdf"), dpi=300, orient=True)
for r in results:
    print(f"page {r.index + 1}: cardinal {r.cardinal}° + skew {r.skew:.2f}°")

Notes & limitations

• The output is rasterized. Without --ocr it is image-only (no selectable text); --ocr restores a searchable layer via OCR.
• jdeskew only resolves tilt within ±--angle-max (default 15°). Gross rotation (sideways/upside-down pages) is handled by the orientation stage, so keep --orient on for arbitrary inputs.
• Tesseract OSD needs a reasonable amount of text to lock onto orientation; on blank or text-sparse pages it falls back to leaving the page as-is (0°).
• --clean uses a local (Sauvola) threshold, which keeps thin faint strokes (e.g. music staff lines) connected far better than a global threshold. It still cannot recover ink that is essentially absent from the scan — very faint lines may stay broken; raising the render --dpi is the best lever there.

Development

uv run ruff check . && uv run ruff format --check .
uv run pyright
uv run pytest                  # unit suite (system binaries mocked)
uv run pytest -m integration   # end-to-end tests against real tesseract/ocrmypdf