rsxdalv/tts-generation-webui

Issue with MusicGen Melody input

Maki9009 opened this issue · 12 comments

It doesnt take Music / Melody input, the CMD says none for that even though i have a song in. it still generates a sound that has nothing to do with the song i input... and its a horrible sound because i assume the melody model needs an input to work properly... there is no error.. it only says that theres none for melody input

Thank you for finding this!
It's fixed in the latest update.

I tested it with the same seed and melody/no melody. And it does make some impact now, although the model doesn't follow it exactly.

Background on the issue: the model "name" changed from melody to facebook/musicgen-melody and so the code kept erasing the melody.

there seems to be another issue now, it does load in a melody.. but i cannot see the progress bar in the CMD... so i dont even know if its actually processing or it froze

Are you using multibanddiffusion? There's no progress for that one. And there's also a small bit of preprocessing for the melody.

Progress gets shown only for the main model.

no im not using multibanddiffusion it set to false, it did generate an audio though... but yeah no progress bar. just tested it again

im talking about the progress counter

Here's how it's shown for me:
image
image
image

yeah i no longer have that n / 1500, it was there before this latest pull i just did

for me it ends with torch.Size([1, 2, 6499544])

That's an unfortunate bug. I tested several times, ran a full update and it still works. Refreshed and reopened the window, also works. I haven't been able to reproduce that.

okay hmm would it be possible to implement a proper progress bar in gradio webui.. or something with a more accurate timing, since the progress bar on there is extremely off

Nevermind, I found it, when doing melody for some reason the progress was disabled.
Yes there can be a proper progress bar as well, but it involves more work.

Pushed as a fix

yeah its working thanks