A neural network is generating faces of people that do not exist and the results are impressive
On the Internet, discerning what is true from what is not is a more complicated task as time goes by. Gone is that era of the fakes and hoaxes of dubious quality that were defended by the poor quality of, for example, the cameras of the moment. However, we are close to making a huge leap in making alternative realities with deep fakes. In video we have already seen worrying realities, but in the photo the new features of the NVIDIA developers are even more disturbing.
The company specializing in graphics cards and increasingly in artificial intelligence began to show face generation results with what they call Generative Adversarial Networks (GAN). However, he still had problems in the faces to be credible to people. What they just showed on video is impressive, the faces look like those of anyone you can meet in real life, without features that are strange in any way.
The GANs offered, according to the NVIDIA research, very little control over the generated images. The new system has a generator that knows how to separate different parts of the images without human supervision. Thus, the different features and parts of a face are taken as much better processed independent variables, which can be combined in any type of face without it being strange.
The responsible team shows in the video how, after the neural networks have been trained, they are able to combine two sources and generate new faces. To do this, they have used 70,000 photographs of human faces from Flickr-Faces-HQ and FFHQ that are available for these investigations (as the Cambridge Analytica data was also available for social sciences). To eliminate statues and faces located in paintings, they have used Amazon's Mechanical Turk platform.
Combined Faces
The system "is not more" than a combination of attributes of different photos (with the possibility of taking precedence over others). In variable A, age, sex or accessories such as glasses are included. In B, which A takes guidance, ethnicity is established. Thus, ** a western child transforms his face into that of the Hindu ethnic group by being merged with that of a woman who has that skin color.
The system works with a collection of styles, where each one controls the effects of a particular part. NVIDIA details three:
The thick part: it takes care of the hair, the shape of the face, the pose, if there are glasses, etc.
Average styles: facial features and eyes.
Fine style: face color.
The video shows that this applies to cars, rooms and animals like cats. However, there the results are not yet as effective. Everything is probably a matter of time.
The company specializing in graphics cards and increasingly in artificial intelligence began to show face generation results with what they call Generative Adversarial Networks (GAN). However, he still had problems in the faces to be credible to people. What they just showed on video is impressive, the faces look like those of anyone you can meet in real life, without features that are strange in any way.
The GANs offered, according to the NVIDIA research, very little control over the generated images. The new system has a generator that knows how to separate different parts of the images without human supervision. Thus, the different features and parts of a face are taken as much better processed independent variables, which can be combined in any type of face without it being strange.
The responsible team shows in the video how, after the neural networks have been trained, they are able to combine two sources and generate new faces. To do this, they have used 70,000 photographs of human faces from Flickr-Faces-HQ and FFHQ that are available for these investigations (as the Cambridge Analytica data was also available for social sciences). To eliminate statues and faces located in paintings, they have used Amazon's Mechanical Turk platform.
Combined Faces
The system "is not more" than a combination of attributes of different photos (with the possibility of taking precedence over others). In variable A, age, sex or accessories such as glasses are included. In B, which A takes guidance, ethnicity is established. Thus, ** a western child transforms his face into that of the Hindu ethnic group by being merged with that of a woman who has that skin color.
The system works with a collection of styles, where each one controls the effects of a particular part. NVIDIA details three:
The thick part: it takes care of the hair, the shape of the face, the pose, if there are glasses, etc.
Average styles: facial features and eyes.
Fine style: face color.
The video shows that this applies to cars, rooms and animals like cats. However, there the results are not yet as effective. Everything is probably a matter of time.
Comments
Post a Comment