NVIDIA uses deep learning to generate extremely realistic looking faces

[vanished]

This is absolutely game changing.  As more of these technologies emerge and they get more capable and easier to use, it will enable amazing things and change the world in incredibly dramatic ways, both for better, and for worse.

[williamcll]

You know what this could be used? Generating pictures of children for expecting parents.

[Jito463]

Did anyone else misread the title as "feces"?  That's some disturbing imagery, I tell you.

[Kamina]

Nvidia needs to generate the perfect waifu for me.

[RejZoR]

On 12/17/2018 at 12:09 PM, leadeater said:

Well that went there fast

That already exists. 3D generated pr0n is actually a thing already. And based on "scientific research" it looks rather realistic with physics and everything.

[GoldenLag]

47 minutes ago, Kamina said:

Nvidia needs to generate the perfect waifu for me.

sorry, but there is allready a perfect waifu. she is known as Komi-san. the empress and god of Komi-nism

[asus killer]

you could either cure cancer or generate random faces, easy choice. Good use of tech

[laminutederire]

Doesn't seem very new though. GAN are mostly hard to train because of saddle points. This seems to me like a less robust workaround to constrain the training. Ideally you'd want the network to create those feature representation on its own as intermediary representations.

[leadeater]

1 hour ago, RejZoR said:

That already exists. 3D generated pr0n is actually a thing already. And based on "scientific research" it looks rather realistic with physics and everything.

That was a joke about the topic turning in to a porn discussion 

[RejZoR]

5 minutes ago, leadeater said:

That was a joke about the topic turning in to a porn discussion 

Doesn't it all eventually? Just give it enough time and topic about harvesters and tractors will turn into pr0n as well

[SADS]

i can't help but feel that this has more downsides than upsides!

skynet will just make a whole bunch of fake accounts, coupled with chatbots... nobody on the internet is real anymore.

 

as long as they don't figure out how to troll as well, then maybe there is still hope.

[kuddlesworth9419]

With today's technology it's entirely possible to fabricate entire events and no one would be the wiser. You would just need to computer generate some event like an assassination attempt or a skirmish or something, use some computer generated voice and this stuff to create believable people involved in such an event and hay presto you have news. Twist is to suit your agenda and they you have tailor made propaganda. And the majority of people won't question it either as it would come from the government. They would say this is the evidence and wouldn't give anymore. 

 

Sounds real tin foil hat like but it makes you wander doesn't it?

[WereCatf]

On 12/17/2018 at 5:06 AM, leadeater said:

Found the most important part, it can do cats.

 

That's pretty cool, though the eyes on some of those cats are pretty wonky.

<content removed >

Now, the more important question is: when can it do smexy ponies?!

 

Edited December 22, 2018 by SansVarnic
Content removed

[Sauron]

On 12/17/2018 at 11:51 AM, colonel_mortis said:

Note that I'm not an expert on GANs or machine learning, and my understanding of the paper is fairly superficial, so I may well be misunderstanding parts of how it works. If you know more, please do correct me.

I'm not sure about this paper in particular (I haven't had the time to read it yet) but generally, in this field, when people mention "features" they don't mean something as specific and large as a nose; a feature is simply something that identifies a specific subset of pixels (or whatever your data is composed of) in as unique a way as possible, in order to be able to align or even just recognize it across a variety of inputs.

 

A nose is probably internally defined as a combination of pixels that produce features with certain values - which is what allows the machine to generate different types of noses.

[colonel_mortis]

20 minutes ago, Sauron said:

I'm not sure about this paper in particular (I haven't had the time to read it yet) but generally, in this field, when people mention "features" they don't mean something as specific and large as a nose; a feature is simply something that identifies a specific subset of pixels (or whatever your data is composed of) in as unique a way as possible, in order to be able to align or even just recognize it across a variety of inputs.

 

A nose is probably internally defined as a combination of pixels that produce features with certain values - which is what allows the machine to generate different types of noses.

One of the novel parts of this paper is, as I understand it, that the features that are used correspond to "styles", which are things like gender, face shape, nose style, etc. The latent space then spans just the different properties of a face, so pretty much everywhere in the space looks realistic.