Researchers advance machine learning to create videos of people from single stills and paintings.
- Samsung researchers improve AI modeling to quickly create realistic (but fake) videos.
- The model uses “single-shot learning” to create videos from a single picture.
- The application of the tech could be in telepresence, video conferencing and gaming.
An AI can now take a single picture of a person’s face and animate it convincingly. This can lead to animating paintings and photos but also add to the mistrust of images and deepfakes online.
The new method from researchers at the Samsung AI Center in Moscow incorporates facial landmarks from a source face into the facial data from the target face to animate it. The target face will do anything done by the source face, which can be any talking head.
What is different here from previous technology that has been developed to achieve this is the fact that rather than needing a lot of data (like video) to analyze, this approach called “single-shot learning” needs just one image of a person’s face. The video that is generated from that can show the face making a range of expressions and speaking with reasonable credibility.
The new tech front loads the process of facial landmark recognition with a large amount of data from a bank of talking head videos. It then trains the model to be very efficient in connecting parts of the target face with the source.
The new AI also uses the Generative Adversarial Network, having two models compete against each other in creating a more “real” result.
Here’s how adversarial learning works:
Where can you use this technology, other than contributing to the epidemic of fake news that is sure to eventually affect personal relationships as much as the national conversations? “Such ability has practical applications for telepresence, including videoconferencing and multi-player games, as well as [the] special effects industry,” Samsung said.
You can read the paper from the Samsung AI Center here.