View Single Post
Old 01-25-2020, 04:19 AM   #7
sai'ke
Human being with feelings
 
sai'ke's Avatar
 
Join Date: Aug 2009
Location: NL
Posts: 847
Default

I think you underestimate how cherry picked the samples are that people typically show.

While some of the advances are indeed exciting, as far as I know we're quite far from having actual controllable performances. As with everything, getting a few impressive cherry-picked samples is "easy" (actually extremely hard), but getting fine control of the process is even more difficult.

I do find some of the wavenet stuff pretty exciting, but if you've ever played with the code they've released, you start to realize that it's not nearly as "ready" as they make it seem. And that you need massive amounts of data to train in these networks. Google is in a bit of a special position here with the massive libraries (*cough* youtube) that they have access to. The problem with most generative NN stuff: you need massive amounts of (often curated) data to train these things in.
__________________
[Tracker Plugin: Thread|Github|Reapack] | [Routing Plugin: Thread|Reapack] | [Filther: Thread|Github|Reapack] | [More JSFX: Thread|Reapack]
sai'ke is online now   Reply With Quote