We're pretty quick to complain about MTL, but maybe this is a newbie redrawer/typsetter trying to figure out how to do things on their own by recreating someone else's work?
I may be being too generous.
Definitely too generous, as if that was the case then they would have done
some proofreading at least. This is straight (obviously) incorrect mtl.
But you do make a good point in that all translators should start releasing their intermediary data, so mtl could actually use it for training and start actually be good.
- original raws
- the text regions identified for ocr (aka text, inside and outside bubbles)
- the ocr results
- the ocr results after human QC
- the translation of the text
- the regions identified for redrawing (aka text outside bubbles)
- the result of redrawing said regions (possibly both without any text, and the region english text is superimposed, depending on how this works).
Then maybe people could manage to actually train a couple proper AI's on some of those steps (image + regions -> identify regions with text; region with text + QC transcribed text -> transcribing text; translation ai's exist already but could train them further to improve in the context of manga; same for drawing out features), and then create a pipeline using them and their results (and simple operations like cropping out the region to hand to next ai).
Large amounts of labeled data is always the holy grail.