Upload forms often cap attachments at 2 MB or 5 MB, and a scanned document blows past that after a few pages. The reason is simple: each page of a scan is a large photo, usually stored at the scanner's full resolution.
TL;DR: Open the PDF with pypdf, downscale every page image, save it as JPEG at a lower quality, and write the file back out.
The script
import sys, os
from pypdf import PdfReader, PdfWriter
src, out = sys.argv[1], sys.argv[2]
MAX_SIDE = 1700 # longest side in pixels after resizing
QUALITY = 60 # JPEG quality
writer = PdfWriter(clone_from=src)
for page in writer.pages:
for img in page.images:
im = img.image
if im.mode not in ("RGB", "L"):
im = im.convert("RGB")
w, h = im.size
if max(w, h) > MAX_SIDE:
s = MAX_SIDE / max(w, h)
im = im.resize((round(w * s), round(h * s)))
img.replace(im, quality=QUALITY)
writer.compress_identical_objects(remove_duplicates=True, remove_unreferenced=True)
writer.write(out)
print(os.path.getsize(src) // 1024, "KB ->", os.path.getsize(out) // 1024, "KB")
What it did on a real file
My test file was a three-page scanned contract, 7,994,634 bytes, with each page stored as a 2481 x 3509 image. After the script, the file was 194,222 bytes, still three pages. I cropped the first page image and the body text was clearly readable at 1202 x 1700 pixels. Tested with pypdf 6.20.1 and Pillow 12.3.0 on Windows.
Tuning
- If the result is still too large, lower
MAX_SIDEorQUALITYa step at a time and look at the page again after each change. - Check small print, stamps and signatures, not just the headline.
- A PDF exported from Word or another editor has few images to shrink, so expect much less change there.
- Keep the original. Re-encoding is lossy, and you cannot get the detail back from the small copy.
Written with AI assistance. The steps were tested by hand before publishing.
Top comments (1)
For text-only scans, JPEG can be a poor tradeoff: it may add artifacts around glyphs, and replacing an efficient CCITT/JBIG2 image can even make it larger. It may be worth checking the source encoding and comparing a grayscale JPEG with a 1-bit lossless option, especially if OCR accuracy matters.