Deepfakes: how Samsung brought the Mona Lisa to life
Controversial technology that was a porn trend is now animating famous images
Researchers in Russia have created a video that shows Leonardo da Vinci’s Mona Lisa uncannily coming to life thanks to artificial intelligence (AI).
Samsung’s AI lab in Moscow produced a research paper that shows how works of art and images of celebrities can be turned into moving images, The Daily Telegraph reports.
The technology used to animate the Mona Lisa is commonly referred to as “deepfake”, the newspaper says. AI takes the facial expressions of a person in a video and merges it with another clip.
But Mona Lisa isn’t the only work to be given the deepfake treatment. TrustedReviews says that the study also brings to life Marilyn Monroe and Salvador Dali, with all three characters “moving, talking and smiling”.
The tech site says Samsung is hoping that the research can help other tech firms develop “practical applications for telepresence”, where technology allows someone to participate in an event that they are unable to physically attend.
What are “deepfakes”?
Simply put, deepfakes are videos that use AI to superimpose people in videos into another clip.
The practice and name come from a controversial pornography trend early last year, The Sun reports. A user on the chat forum Reddit who went by the name of “deepfakes” edited the face of Wonder Woman star Gal Gadot on to the body of a porn actor using “publicly available” software.
Before the Reddit user was outed by Motherboard and banned from the forum, the deepfakes channel amassed more than 15,000 subscribers who had created porn videos that featured the likenesses of Taylor Swift, Scarlett Johansson and Game of Thrones actor Maisie Williams.
Adult video sites such as PornHub have outlawed deepfake porn videos, but companies are starting to explore the technology and what useful applications it may have.
How was the Mona Lisa video created?
Previously, a convincing deepfake video required an AI that was “trained” using an archive of reference material consisting of tens of thousands of video, images and audio files, The Verge says. Generally speaking, the larger the dataset of reference material, the more “eerily accurate the result will be”
Samsung, however, has managed to create an AI algorithm that can turn a single photo or painting into a convincing deepfake video, the tech site says. The algorithm was trained using only 7,000 clips of celebrities, which were then mapped on to a still image to create a video.
What are the ethical issues?
Experts argue that the technology could be used to create false videos of political figures to “fool entire populations”, the BBC says.
For example, a person could create a video showing a politician promoting rivals or policies that they do not support in real life.
The deepfake porn videos that circulated on Reddit last year also raised ethical concerns, as none of them was created with the consent of the celebrities that featured in them.
Speaking to the BBC, Dave Coplin, head of AI consultancy firm The Envisioners, said the technology is “something that could be really problematic unless we have this conversation. Members of the public need to know how easy it is to create convincing fake videos.”