Replaced all deflated images (of a certain type) with jpeg equivalent

[HinTak]

I feel this could be an FAQ but couldn't find a relevant switch/example. What is the mutool convert / pymupdf equivalent of wanting these with ghostscript's pdfwrite, or combinations of?
-dDownsampleColorImages=true
-dDownsampleGrayImages=true
-dDownsampleMonoImages=true

I.e. basically replacing zlib deflated images with their jpeg / ccitt tiff equivalent.

[HinTak]

Actually downsampling isn't exactly what I want - I want to convert all images from zlib deflate to dctdecode/ccittdecode, at its original resolution. This seems to be a useful /frequent request?

[JorjMcKie]

We currently have a - possible meagre - version for that: page.replace_image(xref, <new image spec>) re-inserts an arbitrary new image under the xref of the existing one.
So you could extract a given image, massage it some desired way and insert it again as if it would be new.

[HinTak]

Thanks for the tips . Here is what i come up with now: https://github.com/HinTak/pymupdf-jbig2-extract/blob/main/lossy-convert.py

It is mostly doing want I wanted - just convert RGB flatedecoded images to jpeg. I have two questions though:

• going via PIL.Image.save vs fitz.Pixmap.save results in much better compression, by default even without the optimized=True PIL key (you can roll back a commit or two before to see the diff - I started off with PIL actually, then wanted to remove that dependency and found the size result to be worse - the public repo has simplified/reverted history compared to my private repo). In one file I use for such test, original is 400MB. Going via PIL gives a 75MB output (and down to 72MB with optimized=True), while via fitz.Pixmap.save gives a 200MB pdf. I am quite surprised by this, since it should be compressing a large raw bitmap in both cases, and should just be libjpeg's default, and may even be identical (except for padding at image edges). Why is Pixmap.save so much worse? Or, what settings are not default? I seem to remember Adobe has different ideas about jpeg quality setting from libjpeg's.

• the other thing I like to ask is what's mupdf's equivalent for not updating.the 2nd field in the ID with a random number? qpdf offers a --deterministic-ID . This should probably be a doc.save option.

[HinTak]

Argh, I have my answer to the 2nd question: no_new_id=True.

[HinTak]

Argh, have the answer to my first question also, after extracting the corresponding objects. Upstream mupdf set progressive=True, set dpi=96, disable chroma subsampling, and set quality to 95. Doing all these in PIL I cab get bitwise identical output. Found all of them in upstream mupdf code too, except the last one: upstream default to 90. 95 is a pymupdf setting?

[HinTak]

Yes, 95 is a pymupdf setting:

PyMuPDF/docs/pixmap.rst

Line 326 in d464133

.. method:: save(filename, output=None, jpg_quality=95)

This concludes everything I want to know now. I'll add these as comments, but keep the PIL code.