Anime characters generated with neural networks

It may be a little bit old but there was a thread on /a/ about AI generated "wayfus" , the system used for this generator is based on adversarial neural networks. Even if it does look like it's a serious research project
This generator run on JavaScript via a library called webdnn which can be accelerated with apple webgpu api
make.girls.moe/#/
makegirlsmoe.github.io/main/2017/08/14/news-english.html

Other urls found in this thread:

youtube.com/watch?v=H_NmRuW6848
twitter.com/SFWRedditGifs

Nobody cares.

Using Veeky Forumsence to create the perfect anime waifu.

Brilliant!

It seems to be rather advanced, it's just that Japanese researchers found fun to use a repository of anime style video game characters as a dataset

What else would you use?

I do.

Neural networks are awesome, and anime can look good.

They arr rook the same

>authors are all Chinese

they can't keep getting away with this

this is excellent, technology is well on its way to making anime real

Dont pretend you dont know how to spell that, you loser.

Can someone ELI5 how convolutional neural nets work and what makes them different from conventional multilayer perceptrons? I'm having a hard time wrapping my head around what the "convolution" part actually means. I know they're translation and scale invariant, but that's about it.

Looks like any other RNG generation

What is unique about this in comparison? It's just using preset "forms" right?

Is this any better than other anime generators?

Top tier scientific research right here.

>It's just using preset "forms" right?
I don't think so, it learned all patterns from images itself (i.e. it learned how to identify eyes, hair, mouth, nose, and how they might look, and that an anime character usually has hair, 2 eyes, 0-1 noses, 1 mouth...)

They group parts of the image together in the assumption that pixels next to each other are not totally independent, and calculate weights etc. for those "sub-images"

fuck

Do it work for anime traps?

I guess you can dumb it down like that. at a very basic level its just a very non linear function that takes properties (hair=brown, eyes=red) and returns anime pics which was fitted to pics of anime girls that are labeled with those properties.

>This paper is presented as a Doujinshi in Comiket 92
Kek

MAKE ANIME GREAT AGAIN

Draw a girl - call it a boy.

And sometimes the girls are not even flat.

Kek

Been posted two thousand times.

That's cool, I GUESS. Now what I really want to know is how long until someone downloads all of sad panda and adversarially generates PORN

HAHA OH WOW. Damn, I have a new goal in life now.

> (OP) (You)
>Been posted two thousand times.
Not with this type of generator, it's far more powerful in a way since the input space can be continuous . This technology can be used with another dataset and tags since it's based on training

Most toolsets are open source, there are many tutorials on basic techniques and since computing power in this case is linked to GPU's power it's gonna be available one day for nearly everyone, with the help of Moore's law and vulgarisation.
The best way to speed up this for porn is to make good datasets aviable; well tagged hentai doujins collection anyone?

No, I meant that one. Though I see they did some improvements in the last weeks. There were +10 threads on /a/ in 24 hours about it and quite a few on /g/. Interesting technology indeed.

>>make good datasets aviable; well tagged hentai doujins collection anyone?
You'd be surprised how well hentai is tagged

>the communists will save anime
Marx is smiling in his grave

It isn't.
>still no racial tags, just interracial
>don't differentiate between male and female NTR
>there's no double creampie
>no tags about specific number of people for orgies greater 3, just gangbang

Actually it's more like...
Draw a boy - call it a cat
Funny to see that to this day wizards keep turning people into animals...

:3

>no training/genetic algorithm to tell the AI what I like/dislike about the waifu and let it generate me a better one
discarded

For me, it's Emilia-tan

You've never been to sadpanda. They are mostly extremely well tagged.

youtube.com/watch?v=H_NmRuW6848
is this you?

>"socialism with Chinese characteristics" is communism
Go home, Deng.

Fourth row left is nice

>l*ftism
Fuck off to

Shrenpai I wuv wroo...

...