I am looking at the code, paper and Meta's demos and i have strong feeling that there is something wrong with the way infill is solved.
On all demos they cut the audio and then do actual infill keeping surroundings intact, while here we are use full output of the xt, which expects network to be perfectly trained to synthesize very same audio, which is clearly not expected from the model that is measured to do opposite.
It seems we need either to init y0 with noise + unmasked audio or merge results from solver on each step or run solver on specific segment only.
Hey,
I am looking at the code, paper and Meta's demos and i have strong feeling that there is something wrong with the way infill is solved.
On all demos they cut the audio and then do actual infill keeping surroundings intact, while here we are use full output of the xt, which expects network to be perfectly trained to synthesize very same audio, which is clearly not expected from the model that is measured to do opposite.
It seems we need either to init y0 with noise + unmasked audio or merge results from solver on each step or run solver on specific segment only.