Thank you Ken and Przemek,
formulas are not converted to images in backups: that is done on the fly by the TeX filter, when the page is requested. Images of (already requested) formulas are indeed cached on the server (I don't know for how long), but that cache does not end up in the backup; only the LaTeX code does. What makes the backups huge are ALL the pictures (jpeg's) from ALL the questions in the course. I tried to unzip the backup as suggested, and they are all there, even if I choose to backup one single quiz. I saw the solution put forward by Przemek, it looks a good way out, BUT there seems to be no easy way to pick the jpeg's themselves: when I unzip the backup, they are all inside a directory structure like
file/a6/676e7a09bd5ac3d70aafc7c4b68a63aa501c41cc
with apparently random names (and no extension). There's a lot of them, for one single coures there are 168 such folders with one to 4 files in each.
Is there no other way to expose the original jpeg's? I'd be able to recognize them by name...
Thank you, Giancarlo