Nemo, try to do an "autopsy" of cited IA pdf by pdfimages (xpdf) that recovers raw pdf images into its pages. You'll find that pages are exotically segmented into a full color background, a strange image, and an inverted image of thresholded image (I presume, used as a mask). Just negating the last one, you can get a decent, light BW image of the page. I could build from the last one a decent BW djvu image:
https://it.wikisource.org/wiki/File:Paolina.djvu , but it.source users didn't like the idea
https://it.wikisource.org/wiki/Wikisource:Bar#Pensiero_in_libert.C3.A0_sulle_immagini_delle_pagine
I presume that this complex structure is somewhat similar of djvu background/foreground segmentation into djvu files, and artifacts are similar.
So, pdf images are not only "compressed", but deeply processed and segmented images.
Anyway: IA image viewer doesn't use at all pdf (nor djvu) but uses jpg from jp2 files; so, if you need a djvu similar, for details, to what you see into the IA viewer, you have to download and process jp2 images to build a decent djvu file.
Is something of this complex IA image processing path documented anywhere? I got my conclusions simply by "try and learn" from IA file "necropsy".
Alex