Samsung AI Can Turn Anyone’s Image Into A Fake Video Of That Person Talking

Samsung has taken another step in making it harder to tell if a video features a real person or not. Its latest research can take a normal image and produce a video out of it.

This breakthrough is made by Moscow’s Samsung AI center. They achieved this result by training a deep convolutional network. They fed the system with lots of videos featuring people’s heads when they are talking to teach it to recognize certain facial features.

The research is presented in "Few-Shot Adversarial Learning of Realistic Neural Talking Head Models.” The videos are not as realistic as some deep fake ones surfacing on the Internet lately, but all Samsung’s system needs is one single picture whereas these other videos require tons of them. However, the video’s quality will definitely be improved if it gets more photo input.

The video down below shows you the expected results of the system. With only one picture of Mona Lisa, Marilyn Monroe, Fyodor Dostoevsky, Albert Einstein, and Salvador Dali, Samsung AI was able to produce videos of them talking:

These fake videos cannot fool anyone if they look close enough or an expert. However, if you compare them with previous products, you will see that the technology has improved vastly in just a few years.

This technology can pave the way for anyone having access to it to create a video of a person with one single picture. That coupled with short audio and you can make a video of anyone talking anything. About the setting, there is a tool for Nvidia’s GAN to make the surrounding look real enough.

With the advancement of AI, it will become increasingly harder to tell if a video is real or fake, especially when these powerful tools are available to more people. Hopefully, the system to distinguish a real from a fake video will be developed fast enough.