4 ms·
The sample they used for training are mixed. Unless they can have enough raw, unmixed sample, this depends on how well they "unmix" them.
by j16sdiz 2y ago
The sample they used for training are mixed.
Unless they can have enough raw, unmixed sample, this depends on how well they "unmix" them.
- anigbrowl 2y agoYes...that's the problem. A problem that could be easily avoided by asking existing professionals what matters and what tools they actually want.
- jononor 2y agoMost ML engineers know that many want more fine grained control. But the straight forward way to train such models is incredibly data demanding. The datasets used for whole image generation consist of several billion images. I do not think anyone has compiled any DAW project / stems projects that are anywhere close to this size. So that is a limiting factor right now. But we will find ways to get there, probably a lot of progress over the next 5 years. Maybe even the next 2.