This is essentially OCR (optical cuneiform recognition) which can have specific challenges because the signs are 3 dimensional depressions in clay and present differently depending on the lighting and shadows which can be all over the place.
The output is sort of standard rendering, it looks like they preserve the layout.
Chinese stele's are famously turned 2D through ink rubbing. Wonder if that'd simplify the task and at least provide a more consistent image to work off of (as compared to a randomly illuminated photo)
Or another intermediary would be to train a model to transform photos to rubbing and have a more standard representation to work off of before training to recognize characters
You don't have to do the rubbing physically - if you can 3d-scan the surface you could do it digitally. But, giving the AI less information to work with sounds like a step backwards.
rubbing are much easier to do. While I'm guessing transcriptions are not. So you could provide a lot of training data.
You could then train one model on the rubbing->text and then each research center could make their own photo->rubbing model for their particular photography setup
Direct link to paper which has good illustrations: https://tau-vailab.github.io/ProtoSnap/