[ 3 / biz / cgl / ck / diy / fa / ic / jp / lit / sci / vr / vt ] [ index / top / reports ] [ become a patron ] [ status ]
2023-11: Warosu is now out of extended maintenance.

/ic/ - Artwork/Critique


View post   

File: 1.52 MB, 1433x3276, Dy6nwAGXQAAktQm.jpg [View same] [iqdb] [saucenao] [google]
3805800 No.3805800 [Reply] [Original]

Only a few more months and anime will be a solved problem. Better get on the machine learning wagon now or get left behind

https://twitter.com/gwern/status/1093701790971953152
https://imgur.com/a/zSzwjAY

>> No.3805804
File: 1.26 MB, 2867x1638, DzFBwKGVYAAAt3Z.jpg [View same] [iqdb] [saucenao] [google]
3805804

Just training on Holo from SFW Danbooru 2017 can make great images but also horror. It's the best I've ever seen a GAN do on anime, way better than MGM and you can run it yourself.

>> No.3805806

>>3805800
>a stupid machine can learn quicker than you can
Fuck machines and fuck AI.

>> No.3805878

Where’s the porn GAN

>> No.3806381

>>3805806
In case you didn't figure out, it's just mix and matching drawings from jap artists and you can do that if you understand enough perspective too.

>> No.3806428

>>3806381
Yeah and that takes ages for a typical ningen like me. We don’t got the perfect memory, the perfect coordination, the perfect copy and paste. Never mind absorbing over 5 terabytes of information and categorizing it accordingly.

>> No.3807319

>>3805800
As a computer science undergrad and artist, I'm not sure how I should feel about this

>> No.3807609

Either way, you win.

>> No.3809533
File: 2.97 MB, 512x512, 2019-02-11-stylegan-danbooru2017faces-interpolations-random_grid_404.webm [View same] [iqdb] [saucenao] [google]
3809533

>"One or more high-end NVIDIA GPUs with at least 11GB of DRAM. We recommend NVIDIA DGX-1 with 8 Tesla V100 GPUs."
oh nevermind haha, maybe in a few years

>> No.3809540

>>3809533
>invest into high-end tech
>create a monopoly over anime
>???
>profit

>> No.3809568

>>3805800
WTFFF

>> No.3809571

>>3809533
This is really insightful to the impact small differences have on facial features.

>> No.3809573

>>3805800
I cannot fucking wait until I can generate perfect porn exactly how I want it. I am so excited it's unbelievable.

>> No.3809583
File: 2.98 MB, 384x384, holostyle.webm [View same] [iqdb] [saucenao] [google]
3809583

>>3809571
yeah, it's pretty fascinating. there are 6ish dropbox links in the posts. i never realized how noticeable nose shadowing effects were on illustrations. also it seems like face shapes are much more consistent than i expected.
https://twitter.com/gwern/status/1095131651246575616

>> No.3809593
File: 134 KB, 1400x921, sample-85f04c67e3332aed633e41a0db882cc2.jpg [View same] [iqdb] [saucenao] [google]
3809593

>>3805800
Currently training styleGAN on SHIBUYA rins. I'll post in a day or two if my PC doesn't melt because this is fucking intensive

>> No.3809595

>>3809583
holo winked at me and my heart beat a little faster

>> No.3809608
File: 1001 KB, 500x281, 1500134232480.gif [View same] [iqdb] [saucenao] [google]
3809608

>>3809595

>> No.3809618

>>3809593
What are your specs?

>> No.3809639

>>3805800
this kinda makes me sad. I'm gonna go draw with my tears

>> No.3809665

>>3809639
this. it's somehow very depressing to think what can happen with the insanely rapid growth of all this stuff

>> No.3809666

>>3809665
I fear this more than automation of labor lol

>> No.3809669

>>3809639
Check out the documentary ‘alphago’ on Netflix. It’s about using ai to beat the world champ in Go. Once they won, no one was happy. This is why no one likes nerds

>> No.3809670

>>3809666
well it isn't that different is it?
depending on where it all goes, some seriously impressive tools might end up replacing lots of current jobs; jobs related to art, even, which I thought were almost immune to automation not long ago

>> No.3809832

>>3805800
>A machine can reproduce really boring and unappealing faces which are known for being repitive, samey, and essentially copy pastes of identical assets and features, only slightly smushed around.

Yeah, I'm shakin in my boots. Who would even buy robot art when there's a crowd of mediocre artists who can do it for infinitely cheaper and with 'soul'. Nuff said, goodbye with these threads.

>> No.3810669

>>3809533
Nah 1080s are fine.