2D Machine Learning Graphics

BreezeIndigo

Newbie
Aug 30, 2016
15
16
Sooo...

I trained a neural network to generate graphics for a slave training/breeding kind of game I've sort of had kicking around in the back of my head for a while. The results are...only mildly nightmarish, but the beauty of it is that the network (eventually, when fully trained) can generate thousands of body and face variations and then do the morphing and animating.

See for yourself.
 

ABaaaadIdea

Newbie
Aug 7, 2018
87
181
Fascinating! I *just* started learning ruby/coding in general, so, I'm sure this is well beyond me right now ... but I'm really curious to see how you progress!
 

rainbowpunfz

Newbie
Dec 4, 2017
42
43
StyleGAN2

So the idea would be to learn to map the latent space well enough such that you can morph between different characteristics / poses without hopefully changing other things too much? Interesting and ambitious.

I can see how one would use a GAN to generate fake portraits of in game characters as there is a wide variety of resources out there, but full on nude bodies morphable between some characteristics and poses might form somewhat of a data gathering problem i think? I know someone attempted to synthesize at some point, but GAN's were quite a bit less developed back then.
 

BreezeIndigo

Newbie
Aug 30, 2016
15
16
So the idea would be to learn to map the latent space well enough such that you can morph between different characteristics / poses without hopefully changing other things too much? Interesting and ambitious.
There's a second piece of software called GANSpace that does that for you. Just gotta find the right parameters. Honestly, the bulk of the work is getting a dataset together. All of this stuff is on GitHub, which is rather nice. Honestly, the heavy lifting has been done by others.

I can see how one would use a GAN to generate fake portraits of in game characters as there is a wide variety of resources out there, but full on nude bodies morphable between some characteristics and poses might form somewhat of a data gathering problem i think? I know someone attempted to synthesize at some point, but GAN's were quite a bit less developed back then.
Well, the fake portraits for NPCs idea is where all this started, then I decidedly got pornier.

Also, funny story but vaginas were the second thing I tried. I actually have trained datasets for those. The results were...mixed, but much less horror-show than that guy's. You can transfer learn off of the FFHQ dataset tolerably well, the annoying thing is getting a decent sized set of images from similar angles. I find the StyleGAN2 DiffAugment version works slightly better even than StyleGAN2-ADA for a really small dataset.
 
  • Like
Reactions: rainbowpunfz

rainbowpunfz

Newbie
Dec 4, 2017
42
43
There's a second piece of software called GANSpace that does that for you. Just gotta find the right parameters. Honestly, the bulk of the work is getting a dataset together. All of this stuff is on GitHub, which is rather nice. Honestly, the heavy lifting has been done by others.
Oh that's interesting! I'm not really in the loop w.r.t. the generative side of things. I'd be interested to see where this goes (y)