NVIDIA-Powered Neural Network Produces Freakishly Natural Fake Human Photos (hothardware.com)

MojoKid writes: NVIDIA released a paper recently detailing a new machine learning methodology for generating unique and realistic looking faces using a generative adversarial network (GAN). The result is the ability to artificially render photorealistic human faces of "unprecedented quality." NVIDIA achieves this by using an algorithm that pairs two neural networks -- a generator and a discriminator -- that compete against each other. The generator starts from a low resolution image and builds upon it, while the discriminator assesses the results, sort of like a constant critic, pointing out where things have gone wrong. The GAN is not a new technology, but where NVIDIA differentiates is through the progressive training method it developed. NVIDIA took a database of photographs of famous people and used that to train its system. By working together, the neural networks were able to produce fake images that are nearly indistinguishable from real human photographs, and a little creepy too.

Comments

Not Bad

[mentil]

A few of those example results are a little uncanny valley-ish, but the best are nearly good enough to serve as my dating profile picture. Google Image Search THIS!

Re:Not Bad

[jellomizer]

The real trick is when they are animated.
I remember a back in 2000 where they were showing screen shots of the upcoming final fantasy movie. The screen shots looks like real people without the uncanny valley. However when they started moving and talking then it came to light.

Granted graphics and animation have improved greatly in the past 18 years but I hold my doubts until I can see the rendered images move and interact.

Re:Not Bad

[cellocgw]

Don't be deliberately stupid. The bulk of the profit from movies goes to studio owners and producers. You want them to get even more?
It's like Jim Bouton said of player salaries, "[the players] don't deserve the money, but the owners don't deserve it more."

Do they look creepy?

[Chrisq]

Do they look creepy? They look like many or the retouched "real" photos you see in the media all the time to me!

Curriculum learning

[LetterRip]

This sounds like the standard idea of curriculum learning - you teach NNs via progressively more difficult tasks.

Re:Curriculum learning

[LetterRip]

It isn't in GANs one model, the generative, tries to fool the other, classifier one, by giving it images itself has generated. There's no incremental task switching from easier problem domains to more difficult ones.

I'm familiar with GANs - what it sounded like (and what they did) is add curriculum learning, but they also did it layerwise as is done with autoencoders, (Also they had some other interesting ideas, but that was the crucial bit). In this case the easy is the lower resolution images and the hard is the higher dimensional images.

From their paper

The idea of growing GANs progressively is related to curriculum GANs (Anonymous), where the idea is to attach multiple discriminators that operate on different spatial resolutions to a single generator, and furthermore adjust the balance between resolutions as a function of training time. That work in turn is motivated by Durugkar et al. (2016) who use one generator and multiple discriminators concurrently, and Ghosh et al. (2017) who do the opposite with multiple generators and one discriminator. In contrast to early work on adaptively growing networks, e.g., growing neural gas (Fritzke, 1995) and neuro evolution of augmenting topologies (Stanley & Miikkulainen, 2002) that grow networks greedily, we simply defer the introduction of pre-configured layers. In that sense our approach resembles layer-wise training of autoencoders (Bengio et al., 2007).

Really?

[Anonymous Coward]

"two neural networks -- a generator and a discriminator"

IOW a democrat and a republican. :-)

Training database seems skewed

[swb]

The rendered images look strikingly like actual human photographs, I'll bet they could fool nearly everyone -- you'd have to have a reason to think they were fake.

I'm wondering if their choice of celebrities as the training database somehow skews their results positive versus "ordinary" people. Celebrities almost seem too uniform in terms of facial features and general appearance. It makes me wonder if they tried with ordinary people if the algorithm woudln't produce freaks because it sees odd deviations among normal people.

Re:Training database seems skewed

[BradleyUffner]

The rendered images look strikingly like actual human photographs, I'll bet they could fool nearly everyone -- you'd have to have a reason to think they were fake.

I'm wondering if their choice of celebrities as the training database somehow skews their results positive versus "ordinary" people. Celebrities almost seem too uniform in terms of facial features and general appearance. It makes me wonder if they tried with ordinary people if the algorithm woudln't produce freaks because it sees odd deviations among normal people.

If you look at the full paper, this is capable of so much more than faces. There are dozens of pages of every-day objects they generated, from bedrooms, to wine bottles, to boats, and bicycles. A few of them of some pretty obvious warping and distortions, but the ones that don't look like real objects. It's mind blowing.

Can the criminal system keep up?

[geekmux]

Since photographic evidence is commonly used to convict people of a crime, I can't but help wonder if our legal system will be able to keep up with technology in order to avoid the manipulation that may ultimately condemn an innocent person.

It's quite concerning when the term "indistinguishable" is used to describe technology, as 12 randomly selected citizens can be indistinguishable from a group of morons who are unable to tell the difference between real and fake.

Re:Can the criminal system keep up?

[nealric]

Speaking as a lawyer, I'm afraid you have far too much confidence in the judicial system. People have been convicted based on a lot less than a seemingly perfect photograph and few criminal defendants have the financial wherewithal to hire an expert to contest the veracity of a spoofed photo.

Upscaling application?

[Tx]

You can't get back detail that is missing from a low resolution image, so you can't go e.g. from an SD resolution movie to a 4K one, or at least the result won't look like a movie shot in 4K. Conventional upscaling is basically interpolate-and-sharpen, and it gives only a minor improvement. But while you can't get back the original missing detail, what you could in theory do is generate plausible synthetic detail.

Since this technique seems to involve building up the image through a series of increasing resolutions, I'm wondering if instead of generating a completely synthetic image, you could take a low resolution frame as the starting point, and use similar methods to add plausible synthetic detail. I would have thought that that would actually be a lot easier to generate a good result than if you're trarting from scratch to create a completely synthetic image.

Could it be that our Kazaa-era porn favourites will one day be viewable in 4K quality after all?