Research and improvements in artificial intelligence are evolving and becoming more life-like. Just last week Google declared that they are working on an AI Translator that can translate both your words and the tone of your voice. Now we are seeing another breakthrough made by Samsung’s researchers based in Russia. The company has developed an AI system that brings photos to life by generating a fake video.
In addition, Samsung’s AI Center has published a scientific paper on the model. It’s a new method of applying facial features of a source face to the facial data of a target face. In this way, the target face does what the source face does. The core of the technology in itself isn’t entirely new. It is a result of the synthetic imagery issue that the AI world is trying to crack at the moment.
Another key breakthrough to add realism is from Nvidia. The graphics card firm introduced real-time ray tracing. This technology works by simulating millions of photons, mimicking exactly the way light reflects on different objects in real life. The results add a higher level of realism, making games more immersive and life-like. On the other side, Samsung’s Moscow-based researchers discovered that a video can be created by using only a single image of a person’s face.
So, how does it work?
The model allocates an unevenly huge amount of data of the facial identification process, making it very accurate in distinguishing the features of the target face that correspond to the source. The model, of course, works better when loaded with more data. However, the research shows that it can show good results with one image. This process is called single-shot learning. In this way, you can take a picture of Madonna or Mona Lisa, and make it move and speak:
It also uses a Generative Adversarial Network, that opposes two models against one another. This allows one model to fool the other to make it more “real”. By using the model, the results show a high level of realism that has been designated beforehand. However, the system should reach 90% facial recognition in order to continue. For example, the quality and clarity of some of the fake videos made by the researchers are highly variable. That comes as a result of unknown variables, which the model fills in with gibberish.
The rapid improvements in the AI field mean that any time there is a breakthrough in deepfake photos and video creation, bad actors can begin to mimic it. In this way it can be used for fraud and spreading of fake information, making mass deception easier as deepfakes are harder to disprove. The most famous examples are world known politicians being immitated:
Even though there are still imperfections with Samsung’s new technology, the way it works is still astonishing. Note, that the model works only with the face and upper torso. This means that you couldn’t make Mona Lisa clap her hands or breakdance, yet! One day, it might be possible to do the impossible.