A major known limitation is that it can't do voices, and if you try to get it to recreate a song, it can get stuck in a render->check loop where it starts flip-flopping on details but never gets the score to 100%. I'm guessing that there are some shallow valleys in the score landscape for trying to match the exact sound of a full song.
I put up a showcase page where you can hear some stuff I already made here: https://newsbubbles.github.io/ismail/ to get a feel for some range. I just happen to like brostep so I went that direction immediately.