Samsung's research shows that realistic fake videos are not too far

In an article entitled an unrealistic contradictory learning of realistic models of neural speech researchers at the Samsung AI Center in Moscow and the Skolkovo Institute of Science and Technology revealed the possibility of create realistic fake videos with only a few source images. Although the ability to create compelling deepfakes is not entirely new, the document shows that it is easier and faster than previously imagined.

Currently, to generate a realistic deepfake, you need a large set of image data to form the model. However, the algorithm developed by Samsung researchers can create compelling animated portraits from a small data set, ranging from 1 to 32 images. To do this, they form their model on "historical" facial features, including the eyes, the shape of the mouth, the length and shape of a nasal bridge.

To demonstrate their work, the researchers created living portraits of Mona Lisa, Albert. Einstein, Fyodor Dostoyevsky, Marilyn Monroe and others using only the source image for each. The one-shot results already look impressive, but the realism can be further enhanced by using multiple workouts, as shown in the video below.

The document states that the technology has "practical applications for telepresence, including (videoconferencing) and multiplayer games, as well as the special effects industry. Although companies investing in this space may not have harmful objectives, there is no guarantee that their technology will be used solely for the intended purpose.