Re:There are infinite worlds to explore
by ShamelessC on Monday November 22, 2021 @06:03AM
(#29305107)
Diffusion models are better - but their inference times are orders of magnitude slower.
GAN models are really cool. Don't get me wrong. But they are popular _because_ they are fast at inference. They are still challenging to train (hopefully projected GAN helps with this) and have a very tough time representing highly diverse datasets with lots of data.
> You can actually hook algorithmic art from Houdini to GAN models and bake 3D art for later use.
I've seen some stuff like that but it's not really what I'm talking about. I want fully dynamic, spatially consistent generative worlds.
And that is pretty hard to do at 75 or so FPS in stereo.
You can actually hook algorithmic art from Houdini to GAN models and bake 3D art for later use.
Problem is that GAN models are still not good enough to produce human level realistic images.
It can generate trippy images but these are not very useful for commercial products.
We still need better models or a completely different approach for computer generated realistic art.