Note duplication/corruption
I noticed that at the end of the song there's frequent duplication of a certain fragment which breaks cover generation there because the rhythm and size break too. Since ABC generation seems to be 100% deterministic, here are the two examples:
https://www.youtube.com/watch?v=_VEOjaNvQtI (last chorus, at 03:15)
https://www.youtube.com/watch?v=0FCvzsVlXpQ (at 02:21 and until the end, completely borked; I understand that the tune is audibly messy at that point but after that it's pretty clean and the model worked well until this point)
This doesn't seem to be related to the sliding window trick because both tracks are within the window (300 seconds). Is it a model defect, something with the code, or I'm just being unlucky with these samples? I use ComfyUI and according to Qwen 3.8 the implementation there is identical to the official one. I tried changing the sliding window size, reducing it so this algorithm is actually used (down to 150 seconds with 100 seconds overlap), and this indeed fixes the duplication in the first example at least. But at some window sizes there are similar artifacts in the other places. So my theory is that model's capabilities get noticeably worse at the window end.