This looks impossibly good. I think these samples are cherry-picked and I also think the system is essentially overfit on these datsets and would not generalize

This looks impossibly good. I think these samples are cherry-picked and I also think the system is essentially overfit on these datsets and would not generalize to anything even slightly different. I want to see their failure cases! Lack of failure cases is a red flag. Still, though, it could be useful in its current form, and making a more general system might mostly be a matter of collecting appropriate training data. Impressive work, just needs a more realistic presentation.