A team of researchers with Samsung have found a new way of creating deepfake-style videos based entirely off of a single image.
In a paper published online this week, the team, hailing from Samsung’s AI Center in Moscow, explained how unlike most methods, which attempt to digitally paste one face onto another, their fake videos are created by teaching a computer to understand universally common facial movements.
Referred to as “few-shot learning,” the new algorithm was developed after initially feeding a program thousands of celebrity images and videos in order for it to learn about the human face. Once equipped with such knowledge, the program can then apply what it has learned to a single image it has never seen prior.
The method allows for increasingly realistic videos depending on the number of images. While traditional deepfakes require anywhere from hundreds to thousands of images of a specific individual, the team’s algorithm can use anywhere from one to 32 for comparable results.
A video published on YouTube Tuesday shows the new algorithm in action.
While far from perfect, the video reveals how the technique can be used not only for living people with a wealth of source imagery but even for individuals whose likeness is not as well documented.
In one such example, the team is able to animate the face featured in the Mona Lisa painting. A singular photograph of iconic American actress Marilyn Monroe is also brought to life.
The method represents yet another step forward for fake video creation, which has become increasingly popular and accessible in just the span of a few years.
READ MORE:
- How to spot a deepfake
- Elon Musk is a giggling baby in the latest terrifying deepfake
- This deepfake takes Bill Hader’s Schwarzenegger impression to the next level
H/T Vice
