How can we re-purpose make.girls.moe to generate user tailored porn?

...

Old Thread

The "open mouth" setting seems to cause problems; I've seen very few successful outputs with open mouths.
Also, apparently AI's can't draw hands.

THis is extremely intresting as an artist
I wonder if I could use this to make some pre-generated idea templates
Like using keyord combinations bu already visualised

How did they fucking got this through peer review, out of 20 tries I got 1 image that is somewhat acceptable

When we make a porn version of this we could use a spectrum rating system to kill off the weird looking shit.

And by spectrum I'm not talking about autism kek

>Am I kawaii Onii-chan~~~

Have you put in some weird settings? I'm getting decent images at least 2/3 of the time.

youtube.com/watch?v=ojk9KAbqt1E

AI + Artist = Fuck Yes

Drills seems to be broken, I got stuff like frequently with it, otherwise I didn't use the advanced tab and never hat ribbon hat and glasses on at the same time

...

>Drills seems to be broken
Ouch.
I don't think I've seen the network actually get glasses right either. I suspect the voting has pretty much erased them.

Yeah glasses sometimes cause weird additional structures, but mostly they don't occur at all.

Help her

The love of my life

WHAT DID YOU DO

So this... is the power... of CMU... whoaa...

You can't. You'd have to train a completely new net to do this

If you try a combination of features that isn't represented much in your typical anime girl image database you're going to get weird results because that part of the learned latent manifold is degenerate.

yup. main reason this works as well as it does is because of the (comparatively) clean dataset the authors collected. past attempts (with booru datasets) give much worse results github.com/tdrussell/IllustrationGAN

OP should probably look into
github.com/tkarras/progressive_growing_of_gans
and
tcwang0509.github.io/pix2pixHD/ for more SotA image generation reference

generating a full distribution of porn images from scratch (i.e. random noise inputs) using existing architectures won't work–you could tune it carefully and collect plenty of data and you'd still get horrifying results, because no nn today is generating principled representations of anything from scratch, it's all interpolative

if you had a good annotated-human (e.g. body part segmentation) dataset, that's more promising, because then you can do conditional image generation _from_ those principled, annotated inputs, and then the nn can do realistic image-space rendering without needing to understand anatomy and 3D perspective.

otoh if you want to make a bunch of porn images with the exact same lighting, pose, etc. (and have a dataset for it, like the anime images in mgm) then a pure nn can learn that just fine.

you get the source code of a neural network or build one yourself. Feed data into it. you can repurpose any simple neural network to do this. there was a great youtube vid on a gentleman who repurposed a NN on gihub to mimic human speech, and he used his own voice for the learning.

I am willing to point my nvidia GPUs at this for a while. I have a motherboard with 4 1060s just lying here.

also, check out this version of 3dpd generated by adversarial NNs

hothardware.com/news/nvidia-neural-network-generates-photorealistic-faces-disturbing-results

Advanced mode is fun.

>Get a training data set consisting of every single hentai doujin on one of the big websites
>Generate infinite nonsense doujins