The current 2-up implementation was developed from the British Library's materials - but in those cases the facing pages were already cropped and framed to give a consistent viewing experience, so the problem wasn't seen. I suspect it would have been before long.
What's happening currently The UV code is postitioning the recto tile source on the canvas so that it aligns with the top of the verso image, and adds a small gutter between them (the size of this gutter is configurable, the current value is the result of user testing at the BL). What happens next is down to the OpenSeadragon library [0] that we use to render the deep zoom images. Currently this is what is scaling the images to be the same width. We need to prevent it doing that.
Quick fix - The only metadata the UV has to go on is that the pair of images are the verso of one page and recto of the following page - and even that information is inferred from the sequence rather than explicit in the metadata (but see https://digirati.atlassian.net/bro