Still images come to life in a AI model developed by Samsung at its Moscow research lab, ZDNET reports. With as little as a single frame, speech-like sequences can be generated from reference images, such as a rare photo. Researchers claim with 32 or more reference frames, the results can be highly realistic. A second of video is 32 frames.
Samsung suggests several applications: improved video calls, better player presence in multi-player games, and more believable character special effects in film and television.
Others say it adds to the arsenal for deep fakes, deceptive imagery intended to misrepresent statements or actions by well-known personalities, including politicians and authority figures.
The system works by extracting features from the source picture and mapping them to expressions and movements drawn from a huge database of images. It uses as AI technology known as generative adversarial networks, often more simply called GANS. The research paper calls the approach ‘few shot adversarial learning.’
- A great illustration of the two-edge sword of some AI advances: the same technology benefiting ‘video presence’ for consumer products at the same time makes deepfakes easier to create
- Imagine what Mona Lisa, Einstein, or others may have been thinking… look for companies to do just that.
- New AI can create fake videos of people from a single picture
BIG THINK | May 24, 2019 | by Paul Ratner | ‘The target face will do anything done by the source face, which can be any talking head.’
- ‘Few-Shot Adversarial Learning of Realistic Neural Talking Head Models’ | Original research paper