NVIDIA in collaboration with Aalto University (Finland) created an AI that creates fake celebrity photos that look so real

[captain_to_fire]

Source: NVIDIA via CNET

 

I can't upload the actual paper because of the attachment limitation to 20 MB of LTT so just click on the NVIDIA link above to read the actual paper which is 28.7 MB. 

Quote

Progressive Growing of GANs for Improved Quality, Stability, and Variation

Tero Karas (NVIDIA), Timo Aila (NVIDIA), Samuoi Laine (NVIDIA), Jaakko Lehtinen (NVIDIA and Aalto University, Finland)

Under review as a conference paper at 6th International Conference on Learning Representations

 

ABSTRACT

We describe a new training methodology for generative adversarial networks. The key idea is to grow both the generator and discriminator progressively: starting from a low resolution, we add new layers that model increasingly fine details as training progresses. This both speeds the training up and greatly stabilizes it, allowing us to produce images of unprecedented quality, e.g., CelebA images at 1024². We also propose a simple way to increase the variation in generated images, and achieve a record inception score of 8.80 in unsupervised CIFAR10. Additionally, we describe several implementation details that are important for discouraging unhealthy competition between the generator and discriminator. Finally, we suggest a new metric for evaluating GAN results, both in terms of image quality and variation. As an additional contribution, we construct a higher-quality version of the CelebA dataset.

I think the likes of TMZ or other famous celebrity gossip website would love this. 

Quote

The graphics chip maker published a study last week on its use of "generative adversarial networks," or a kind of artificial intelligence that employs two neural networks, to create photo-realistic images of fake celebrities. See those photos above? Those aren't real people. The process begins with a low-resolution image, and the two neural networks work to lay out more high-definition features until you get to those finished results. The company used the CelebA database of stock celebrity images to create different mash-ups. 

From the paper:

Quote

We use generator and discriminator networks that are mirror images of each other and always grow in synchrony. All existing layers in both networks remain trainable throughout the training process. When new layers are added to the networks, we fade them in smoothly, as illustrated in Figure 2. This avoids sudden shocks to the already well-trained, smaller-resolution layers. Appendix A describes structure of the generator and discriminator in detail, along with other training parameters.

 

We observe that the progressive training has several benefits. Early on, the generation of smaller images is substantially more stable because there is less class information and fewer modes (Odena et al., 2017). By increasing the resolution little by little we are continuously asking a much simpler question compared to the end goal of discovering a mapping from latent vectors to e.g. 10242 images. This approach has conceptual similarly to recent work by Chen & Koltun (2017). In practice it stabilizes the training sufficiently for us to reliably synthesize megapixel-scale images using WGAN-GP loss (Gulrajani et al., 2017) and even LSGAN loss (Mao et al., 2016b).

 

Another benefit is the reduced training time. With progressively growing GANs most of the iterations are done at lower resolutions, and comparable result quality is often obtained up to 2–6 times faster, depending on the final output resolution.

Just so you know, those pictures above aren't pictures of real people They're just computer generated. Although in Figure 5, one of the pictures kinda looks like Beyonce. The researchers concluded that it still needs improvement.

Quote

While the quality of our results is generally high compared to earlier work on GANs, and the training is stable in large resolutions, there is a long way to true photorealism. Semantic sensibility and understanding dataset-dependent constraints, such as certain objects being straight rather than curved, leaves a lot to be desired. There is also room for improvement in the micro-structure of the images. That said, we feel that convincing realism may now be within reach, especially in CELEBA-HQ.

To be honest I only understood a few of it but from the looks of it, it can be used by the likes of Apple to train Face ID in their neural networks in order to spot impostors or fake news websites can use it to create mass hysteria.

[ZestyJalapeno]

It's going to be used for pr0n, guaranteed

[Zodiark1593]

This would be perfect for my Satire paper. 

[Likwid]

How to get accused for something you never did

#Tinfoil~Hat~Intensifies

[dizmo]

Good to see all that brain power going to something so pointless, and potentially extremely dangerous. 

[PlayStation 2]

1980: I bet we'll have flying cars in 2017

2017: 

[Guest]

Its cool to see that neural networks are improving, but the results still look unnatural. (Still a great deal better than an amateur Photoshop at least)

[captain_to_fire]

56 minutes ago, tjcater said:

Its cool to see that neural networks are improving, but the results still look unnatural. (Still a great deal better than an amateur Photoshop at least)

The one that looked like Beyonce looks sort of real though 

[Guest]

1 minute ago, hey_yo_ said:

The one that looked like Beyonce looks sort of real though 

Won't deny that, especially at a quick glance, but most of these photos just give me that uncanny valley vibe after looking at them long enough (Eg 10-15 seconds).

[Maslofski]

one step closer to endless really good computer generated movies, cant wait

[laminutederire]

1 hour ago, tjcater said:

Won't deny that, especially at a quick glance, but most of these photos just give me that uncanny valley vibe after looking at them long enough (Eg 10-15 seconds).

Those will probably be kings to render faster CG scenes, not so much pictures of people. Because they can get more knowledge this way and make something better looking

[shadowbyte]

wow

[Prysin]

3 hours ago, Dogeystyle said:

It's going to be used for pr0n, guaranteed

rule 34, rule 35

[LukaH]

this is why AI will kill us.

[SpaceGhostC2C]

3 hours ago, Dan Castellaneta said:

1980: I bet we'll have flying cars in 2017

2017: 

2017: well, we can have a picture of you on a flying car.

[NumLock21]

My school never taught us the RGB formula...?

[captain_to_fire]

12 hours ago, Likwid said:

How to get accused for something you never did

#Tinfoil~Hat~Intensifies

But how can they accuse someone when the picture isn't from a real person?

[NumLock21]

1 minute ago, hey_yo_ said:

But how can they accuse someone when the picture isn't from a real person?

Out of the 7 billion people on this planet, there might be a chance a few of these computer generated people is real.

[captain_to_fire]

Just now, NumLock21 said:

Out of the 7 billion people on this planet, there might be a chance a few of these computer generated people is real.

True as they made an image looking so close to Beyonce but a picture of a person looking like a vampire is probably fake anyway

[NumLock21]

4 minutes ago, hey_yo_ said:

True as they made an image looking so close to Beyonce but a picture of a person looking like a vampire is probably fake anyway