Actually, I would rather have a tool that pulls apart djvu files as they are uploaded; keeping the text in WS and the pics in Commons
This is very interesting since abbyy.xml files contain both a full detail (character by character) detail of text mapping & format, and coordinates of any not-textual content (illustrations) of the scanned page. Using appropriately such data, it would be possible to extract automatically illustrations and other graphical elements of pages. nevertheless, I saw that such "self-cropping" of illustration sometimes fails, and often is confused by some unusual format of illustrations/graphical element, so that many "illustrations" are nonsense or have to be cropped again. Unluckily, djvu files have no such "illustration coordinates" inside.