Artificial intelligence or AI can be a useful tool to help take the burden off our shoulders when it comes to trivial and mundane tasks. Companies like Google have used AI and machine learning to even emulate people for simple tasks such as booking a salon appointment, with products such as Google Duplex. AI algorithms have also been used to animate 2D objects such as talking heads of people, but this typically requires a huge dataset of samples to train the neural network to adapt to the objects or in this case, the person's behaviour and mannerisms. This process just got a lot simpler thanks to Samsung AI Center's Moscow chapter, which has published a paper which shows that a highly realistic talking head model can be obtained by feeding the algorithm with as little as a single sample photo.
The research paper[1], titled 'Few-Shot Adversarial Learning of Realistic Neural Talking Head Models', from the Skolkovo Institute of Science and Technology in Moscow, found its way[2] to a sub-reddit along with a demo video which shows the AI in action. The algorithm uses a technique of using "landmarks," which is a skeletal framework of a persons face learnt by crawling through a large dataset of videos, which is used as the backbone for the synthesised face. The realism of the synthesised face depends on how many source images you feed the algorithm.
The paper[3] shows the difference between models trained with 32 samples, eight samples, and one sample. The output from just one sample is obviously has a few artefacts around the face, helping you spot that it's not an actual video but an AI-generated one. But the fact we can even get such output with a...