manga-page-splitter

Split scanned manga double-page spreads into single pages, in right-to-left reading order.

Scanned manga is usually distributed as one image per printed spread: two pages side by side, separated by the fold of the book. That is fine for reading, but it is awkward for anything that has to work page by page — e-ink readers, translation tooling, OCR, re-binding a volume into a single PDF with a sane page count. This package does the boring part: it finds the fold, cuts the spread in two, and hands you the pages in the order they are meant to be read.

Installation

pip install manga-page-splitter

The only runtime dependency is Pillow.

Quick start

from manga_page_splitter import split_file

# Writes page_001.png and page_002.png into ./pages
written = split_file("spread.png", "pages")
print(written)

From the command line:

manga-page-splitter spread.png -o pages

Reading order

Japanese manga is read from the right page to the left one, so by default the right half of the spread is emitted first. For comics that read left-to-right, pass right_to_left=False or --ltr:

pages = split_spread(spread, gutter=400, right_to_left=False)

manga-page-splitter spread.png --ltr

Finding the gutter

gutter is the x coordinate the spread is cut at. You can supply it yourself, but you usually do not have to: detect_gutter scans a band around the horizontal centre of the image and returns the lightest column in it. On a real scan the fold is printed with no artwork across it, so it is almost always the cleanest vertical strip near the middle.

from manga_page_splitter import detect_gutter

x = detect_gutter("spread.png")
if x is None:
    print("no clean fold found — pass an explicit gutter")
else:
    print(f"cutting at x={x}")

The search is configurable:

Raise min_ink_ratio for noisy scans, lower it to demand a genuinely clean fold. When nothing in the band qualifies, detect_gutter returns None and split_spread raises GutterNotFoundError rather than guessing — a wrong cut silently corrupts a whole chapter, so failing loudly is the better default. Pass gutter=<x> to force a position.

To check a folder of scans before committing to a batch:

manga-page-splitter scans/*.png --detect-only

Trimming the margins

Scans usually carry a white or grey margin, and a few millimetres of the facing page. trim=True (or --trim) crops each page to the content box, using the colour of the four corners as the background:

pages = split_spread("spread.png", trim=True, padding=4)

padding keeps a few pixels of breathing room around the artwork, which is usually what you want if the pages are headed for a reader that adds its own margins.

API

Every function accepts either a PIL.Image.Image or a path. split_result returns a frozen SplitResult with pages, gutter, size, detected and a count property.

Command line reference

manga-page-splitter IMAGE [IMAGE ...] [-o DIR] [-g X] [--ltr] [--trim]
                    [--padding N] [--prefix STEM] [--format FMT]
                    [--quality N] [--batch] [--detect-only] [-q]

Exit codes: 0 on success, 1 for I/O or argument errors, 2 when no gutter could be found.

Batch use

# Each scan gets its own folder: pages/chapter01/page_001.png, ...
manga-page-splitter scans/*.png -o pages --batch --trim

In Python, split_many does the same and returns one list of paths per source file.

Notes and limits

• Only two-page spreads are handled. A single page passed in will either be cut in the wrong place or rejected, depending on how much ink sits in the centre band.
• The gutter search assumes the fold is roughly centred. Heavily cropped scans where one page takes two thirds of the width need an explicit gutter.
• Colour profiles are preserved; nothing is re-encoded unless you ask for a lossy --format.
• Pages are written in reading order, numbered from 001, so the file names themselves sort correctly.

Related

Cutting a spread in two is the first step of a scanlation workflow; the harder half is the text. Mee Manga Translator takes the separated pages and translates the dialogue while keeping the original layout, which pairs well with the --trim output above.

License

MIT. See LICENSE.