@genl Oh crap I have specified a wrong parameter. Though it doesn't quite matter. Upon closer inspection, most if not all greyscale drawings are somewhat upscaled while the texts are in native resolution... Seriously KTCOM this is the best you can bake? lol
Concerning the size, the images extracted are optimized at quality 90 with MozJPEG. DSSIM gives an average of 0.000665 (0 implies two images are identical) for the whole collection which suggests many of the extra bytes are in fact artifacts (as it is upscaled Orz).