Trying vevo2 svc with style preserved. If I use an audio longer than 50 seconds the output is garbage. If using a 40-50 sec vocals, the output is a bit better but still not good enough and after exactly 30 seconds, garbage again. Trying with 10 seconds audio produces much better output.
I think in order for the model to work for longer audio it is mandatory to have sliding window with an overlap. Working in chunks is the only way, at least for me. Is such implementation possible?
I have some problems with style changing and asking for input text but I guess that will be a separate topic.
Ps Thanks for fixing the out of memory problem.
Trying vevo2 svc with style preserved. If I use an audio longer than 50 seconds the output is garbage. If using a 40-50 sec vocals, the output is a bit better but still not good enough and after exactly 30 seconds, garbage again. Trying with 10 seconds audio produces much better output.
I think in order for the model to work for longer audio it is mandatory to have sliding window with an overlap. Working in chunks is the only way, at least for me. Is such implementation possible?
I have some problems with style changing and asking for input text but I guess that will be a separate topic.
Ps Thanks for fixing the out of memory problem.