Image generation from text

Why is there a background if its cg??
Seems fak4

That's not how GANs work. You need hundreds of thousands (and possibly even millions) of example images in order to generate fakes which look alike.

Watch this video to understand more: https://channel9.msdn.com/Events/Neural-Information-Processing-Systems-Conference/Neural-Information-Processing-Systems-Conference-NIPS-2016/Generative-Adversarial-Networks

>>59185058
It's not cg. I mean, it is, but automatically generated by a neural net based on the training set, which is composed by, usually, real images (= with backgrounds)

>>59185269
>using a neural net to generate images
>results in an actual image (not cg)
anon, i have bad news...

fake custom porn is the future
the only way is you imagination right now...

>>59185099
>Hundreds of thousands
I'm seriously thinking this could be possible, since internet is full of porn

>>59185367
Anon, you should know something about this method before judging this fake
http://blog.aylien.com/introduction-generative-adversarial-networks-code-tensorflow/

>>59185720
He saying it doesn't "generate images"
What you have is advanced google image search

>>59185918
I understand what he says. The point is GANs actually generate images, and those images are generated automatically. It isn't in any way related to research algorithms.

>>59186000
>>59185639
first learn how to distinguish a flower from grass and other distinct imagery you troll

https://open_nsfw.gitlab.io/
How about things that look like porn but aren't?

>>59186000
Are those images rendered or loaded?
You have advanced Google image search, don't be one of those chucklefucks praising AI for EVERYTHING when you see it used to look up an image

>>59186191
>troll
Yeah just call everyone trolls for literally no reason, that doesn't make you look like an old school php forum dropout at all

>>59186727
Haha good meme
Rekt
Btfo
Pajeet
Housefire

>>59186727
Your entire knowledge of neural nets consist of reading a blog and a random pdf last night
You speak like 3rd world terrorist
You expect anybody to take you seriously

>>59186725
For what I know (not much actually) the GAN is a density prediction model, so for each pixel you have a predicted color value based on previous inputs. I don't really know if this can be called 'rendering', but basically what you have is a first image created with the stage 1 GAN specifically trained to generate low res images using text as input. Generate means generate, so it's not google image search. And as you can see in this pic, stage I often results in a weird looking image with some sort of similarity with what is written in the input text. And they are of course not real pics, they are generated.
After this first stage pic is generated, another algorithm is applied to augment the resolution of the picture.

>>59184943
>Could someone use the same logic of this algorithm to generate fake custom porn?
>This flower looks like a vagina.

>>59186909
>generate means generate
>Google Dream means it actually Dreams

>>59186988
kek

>>59186909
So this will operate independently, creatively, without a giant databank?
Would interesting if I didn't need a supercomputer to do this (unless it just matches images to words like captcha)

>>59187007
After the training this shouldn't take much computational efforts

>>59187062
But does it require a databank of images?

>>59187069
Not after the training, like every neural network model i guess