Raw jbig2 data (instead of auto-converted png) wanted from 1.25.2+?

[HinTak]

A fair number of pdf's I encountered consists of a yellowed paper colored image plus a 1-bit mask where the mask is the interesting thing, so I wrote a script in 2024 to just extract the jbig2 masks losslessly into a (smaller and more readable) pdf:
https://github.com/HinTak/pymupdf-jbig2-extract/blob/main/extract-all-mask.py

It worked in pymupdf 1.23.x then. Recently I found that it breaks against current pymupdf (1.26/1.27). I have tracked down the breakage to be between 1.25.1 and 1.25.2 .

In 1.25.1, the two images are of type jpx and jb2, but in 1.25.2, the latter seems to be auto-converted to png, and I am no longer getting the raw jb2 data stream. I see a couple of issues here where auto-conversion to png within 1.25.1->1.25.2 was implemented... my question is, how do I get the old behavior of extracting raw jb2 mask data, rather than a converted png mask in the new pymupdf?

[JorjMcKie]

I have no immediate solution, but you can extract any stream data directly from a PDF via xref_stream_raw. This is as-is, without decompression. Then decompress this as desired/required and save it as an image in some desired format, potentially using Pillow.

Another idea is making a Pixmap from an image xref Pixmap(doc, xref). Then manipulate the pixmap's color and alpha channels in the desired way. Here, numpy can be a great helper.

[HinTak]

It has been two years since I wrote that code, but the part which seems to have changed behavior between 1.25.1 and 1.25.2 is this:

...
il = page.get_images()
    if (print_once):
        print(len(il))
    for img in il:
        xref = img[1]
...

I am still getting exactly two images per page, but the 2nd member, a mask, has changed from jb2 to png. I guess I might need a get_images(do_not_convert), or do a different look up in xref = img[1]['raw']. I'll have a look around what changed between 1.25.1 and 1.25.2 .

[HinTak]

FWIW, the relevant change in that time frame seems to be this #4186 (comment)

[HinTak]

Thanks - I have updated the code to use xref_stream_raw . I timed it too - it seems that 1.28 is faster than 1.25 by about 1.6x somehow (plus another 1.4x from python 3.13 to 3.14).