From 13 GB to 2.7 GB: choosing format and resolution for scanned manga

On my home server I run Manga, my own manga reader: a library with covers, per-person reading progress and a phone app. Almost everything I uploaded went in without trouble, except one series: scanned PNGs of about 800 MB per volume, more than 3 MB per page.

On a computer you don’t notice. On a phone, over mobile data, every page took a while to show up and a full volume ate a real chunk of the plan. It had to get lighter without anyone noticing while reading.

What was in those files

Before touching anything I looked at what I had:

• Grayscale pages, from 1840×3200 to 2400×4164 pixels.
• Saved as PNG, a lossless format meant for graphics, not for photos of scanned paper.
• About 200 pages per volume, 16 volumes: 13 GB in total.

My first idea was to recompress the PNGs without changing anything (losslessly). I tried it and it only saved between 26 % and 53 %: better than nothing, but still huge for reading on a phone.

The comparison, on real pages

Instead of picking a format “because it’s the modern one”, I wrote a script that took real pages from the series and converted them to three formats at three resolutions (maximum height), measuring the size and time of each:

Format and height	Size per page	Versus the original
Original (PNG)	3273 KB	100 %
JPEG 3200 px	1633 KB	50 %
JPEG 2400 px	1010 KB	31 %
WebP 2400 px	686 KB	21 %
AVIF 2400 px	508 KB	16 %
JPEG 2000 px	732 KB	22 %
WebP 2000 px	503 KB	15 %
AVIF 2000 px	378 KB	12 %

Looking only at the table, the answer seemed obvious: AVIF at 2000 px, almost nine times lighter. But a number doesn’t tell you how it looks.

The deciding detail: screentone

Printed manga doesn’t have real grays. Shadows and backgrounds are made with screentone: a grid of tiny black dots that the eye blends into gray from a distance. It’s part of the drawing, and it’s the first thing to suffer when you shrink an image.

I opened the same pages on a phone and zoomed in:

• At 2000 px, the screentone was no longer dots: it had turned into a blurry gray. Reading straight through you barely notice, but with zoom (which you do all the time on a phone) it looked washed out.
• At 2400 px, the dots were still there. And since that height already more than covers any phone screen, more resolution added nothing visible.

So the resolution was set at 2400 px: the lowest point where the screentone survives.

WebP instead of AVIF

At 2400 px, AVIF was 26 % lighter than WebP. I still chose WebP at quality 85, for a few reasons:

• WebP has been supported for years in browsers, readers and comic apps; AVIF is newer and its support is still more uneven.
• AVIF is more expensive to decode, and that shows when flipping pages quickly on a mid-range phone.
• It was also somewhat slower to encode in my tests.
• Across the whole series, the estimated difference was about 700 MB. A reader that runs smoothly everywhere was worth more than that last bit of savings.

The script

The conversion is done by a Python script with Pillow, running in parallel to use every core:

• It keeps the page in grayscale if it already was: saving it in color would mean paying for channels it doesn’t use.
• If the page is taller than 2400 px, it scales it down with a high-quality filter (Lanczos), keeping the aspect ratio.
• If the WebP version is heavier than the original (it happens with nearly blank pages), it keeps the original.
• It writes a new CBZ and never touches the original.

Then a second script compares every volume with its original: same number of pages, and every one of them opens. It checked all 3,280 pages without a single error.

The result

The series went from 13 GB to 2.7 GB (−79 %), about 840 KB per page. Reading on a phone I can’t tell it apart from the original, and pages show up instantly. The originals are still stored separately in case I ever need them.

The rest of the library didn’t need recompressing: the PDFs I had were, inside, one image per page, so the upload script extracts those images as they are, without re-encoding them. The 72 volumes that came as PDFs were converted with no quality loss at all.

What I’m taking away

1. Measure with real data, not internet examples. “WebP is X % smaller” figures don’t help if your images are grayscale scans full of screentone.
2. The smallest number doesn’t always win. AVIF at 2000 px won on the table and lost where it matters: while reading.
3. Look for the threshold, not the extreme. What helped was finding the exact point where quality breaks (between 2000 and 2400 px) and staying just above it.
4. Never recompress without a way back. Keeping the originals and an automatic check made the change safe.