How AI inserts Nicolas Cage into movies and making porn with celebrities

Now you can not believe politicians who say something on YouTube videos, and "plums" of sex video stars. Reddit users taught AI to replace faces with videos, and in recent weeks it has exploded so much that no celebrity has been left untouched. Most of all went to Nic Cage, who, with his expressiveness, had long been a gold mine for memes creators. Thanks to artificial intelligence and machine learning opportunities, this has reached a new level, so now people are also looking at technology in Hollywood. The effects are sometimes better than the last blockbusters. Here are some users who have other interests ...

How did we come to this

The first acquaintance with the idea of ​​"inserting" a person inside the video for many was the film "Forrest Gump", filmed in 1994. With a budget of $ 55 million, the character Tom Hanks turned out to be “embedded” in several famous scenes of American newsreels. But the effect was not realistic to name. Even an untrained user could notice that something was wrong: too Forrest stood out from the background. The best specialists in Hollywood could not imperceptibly insert the finished Tom Hanks into the finished chronicle.

Not very convincing?

Everything has come to a new level literally in the last few years - with the development of artificial intelligence and machine learning algorithms. In 2015, Facebook began to test DeepFace technology among users. The social network gathered the largest photo library in the world, and she wanted to tie as many people as possible into it. Even those that the users themselves did not mark. DeepFace learned to determine whether two persons in different photographs belong to the same person, with an accuracy of 97.25%. According to Facebook, people answered the same test correctly in 97.53% of cases, that is, they were only a quarter of a percent more successful than the system.

Work DeepFace

Knowing who's photos are on Facebook is very important: so she can sell personalized ads even more successfully, and trade information about her users. As a result, DeepFace technology began to work for the “Mordoknigi” almost everywhere, except for the European Union, where it was obstructed by data privacy laws.

In 2016, users began to play with the technology of searching and replacing faces. Filter Face Swap in Snapchat, changing faces in places, became a mega-hit. Daily video with him sent a few hundred million times - and continue to send so far. MSQRD for iOS and Android can also cope with similar tasks (which also bought Facebook after 3 months of creation), and others.

Face Swap Live, another application

Finally, in 2016, developers from Stanford University presented their work (by the way, Snapchat is also former Stanford students). Their technology is called Face2Face and allows you to change facial expressions for video in real time using the most common webcam. The work is extremely interesting (you can read in PDF ), and the results are simply amazing. The hero of any video can “transplant” his facial expressions and speech. The webcam scans your face, and the program matches it with the face of the character, and imposes on it all your expressions.

In the demo video, developers make Bush, Trump, and Putin wry as they please. The technology works in RGB, is able to render not only the lips and cheeks, but also the inside of the character's mouth, and does not use the “geometric proxy teeth arrangement”, which is a special pride for creators. The result is a result that is sometimes difficult to distinguish from reality. Having such a program, any politician can be forced to say anything. “Video evidence”, it turns out, can no longer be trusted, and will stop taking pictures from a security camera or video from a smartphone as evidence in court in a few years. Thanks, Stanford!

Entry to a new level

Then the fun begins. In December 2017, the RedFit user DeepFakes (the obvious allusion to DeepFace from Facebook) enters the arena. With its algorithm that allows you to create realistic videos with the replacement of a person using neural networks and machine learning. With open source available to other enthusiasts ( link to GitHub ). From his home PC, he made the first "masterpiece": porn with Gal Gadot, "Wonder Woman", in the title role.

We used libraries for deep learning Keras and TensorFlow, and open machine learning algorithms. DeepFakes says its algorithm is similar to the one that Nvidia uses in video-changing technology or weather conditions introduced a little earlier in December. First, the system "trains" in the pictures on Google and videos on YouTube. In the process of learning, the AI ​​tries to compare fragments of the original portraits with the video to get the most realistic video. That was how a porn movie with Gal Gadot was created: the car was “fed” with its photos from Google and videos of various interviews on YouTube. And when she figured out what the actress was, the model Gal Gadot was put on a clip with a porn actress of similar proportions.

DeepFakes explains the work of his creation in an interview with Motherboard:

I take hundreds of photos, and generate millions of distorted images from them to train the network. When she is ready and knows what kind of person I need to make, it’s enough to feed the networks of someone else’s face. The network thinks it is another distorted image, and tries to make it look like the face with which she trained.

On Gal Gadot DeepFakes did not stop. To show the possibilities of his algorithm, he made porn with Maisie Williams (Arya from “Game of Thrones”), Scarlett Johansson, Taylor Swift and others.

But everything took off when another user, Deepfakeapp, a month later created FakeApp , so that everyone could mess around the way they want, even without Python knowledge. January 31, the program has reached 100,000 downloads.

FakeApp interface

Creating data for training the neural network, training the model, creating fakes - with the new program it became available without the need to install Python and without knowledge of syntactic analysis.

Immediately send memes. One of the most innocuous among them was Nicolas Cage, whom they began to insert wherever possible. Firstly, the face is characteristic, secondly, it is funny, thirdly, is it true that Nick Cage is not given the best roles?

Cage lois


Cage in "Terminator 2"

According to Deepfakes, it takes an average of 8-12 hours to create one short video from the program. An expensive video card can handle this task pretty quickly, and a budget PC takes a couple of days to create a video.

Cage Superman

The only obstacle for the algorithm may be a lack of source materials, but in the Internet era there will be no problems with this. So, only from 2015 to 2016, netizens have published more than 24 billion selfies.

Two Cage

Also, the creators of AI-fakes made sabreddit with their "creations" (carefully, 18+!). He exploded in popularity, in a few weeks he had 56 thousand users (update: already 65 thousand, in three hours). Each celebrity, which at least someone wanted to see naked, is presented in ten versions. It got to the point that the Gfycat service forbade placing gifs with fake porn because they scored the entire top views.

Cage - Indiana Jones

Three days ago, the channel Deepfakes AI with porn celebrities was blocked Discord. Reddit, on which the idea was born, is silent for now. Well, the video has so far (probably temporarily) found refuge on Pornhub (again, strictly 18+).

Katy Perry

Some fake porn already successfully issued for this. Movie with Emma Watson, the host of the shower, picked up the site Celebjihad, posting the "leaks" of naked celebrities. Considering how many photos are widely available on the Internet for most users of social networks, there are fears that porn can now be done with absolutely anyone. This already has a term - "porno-revenge." If an unlucky suitor or a former boyfriend wants to ruin the life of his passion, he just has to turn to a single programmer with a powerful PC.

Daisy Ridley from Star Wars

Wired reports that the creator of FakeApp can be held accountable if it can be proved that the technology is used to create "unmatched porn." But there was no precedent for the case, and obviously, as soon as a celebrity sues, the traffic to his / her videos will increase many times. Also the legal problem is that the body in the video does not belong to the person filing a lawsuit, and it will be difficult to classify as a breach of confidentiality. “You cannot sue someone for revealing the intimate details of your life if it is not you whom they reveal,” writes Wired.

Many people on the network call the work of FakeApp users "digital rape". But the user from Reddit Gravity_Horse protects their experiments:

What we are doing in this community is not malicious. On the contrary. We are drawing with a revolutionary, experimental technology, which may well build the future of media and creative design. If someone wants to make a fake to harm a person, there is Photoshop. It is much simpler and exists for many years. In addition, the faster this technology will become mainstream, the less effective will be pornomest. Even real scenes shot at home by a former boyfriend can be dismissed by saying that it is fake.

At the same time, it is expected that the technology will be used in Hollywood. For example, the neural network in FakeApp in one day and at a cost of $ 0 made the young princess Leia more qualitatively than the team of experts in “Rogue One: Star Wars” with a budget of $ 200 million. The technology can have a great future in the film industry. As usual, porn is the engine of progress.

Attempts to replace old heroes with computer graphics from Disney and LucasFilm have so far been so-so (still from the movie “Rogue One”)

PS One of the celebrity media elements is real. Can you understand without Google?