The lines separating deepfakes from real people are becoming increasingly blurry. MyHeritage, a genealogy platform, released Deep Nostalgia™, a new feature that uses GANs to animate faces in still photos. Licensed from an Israeli Deep Learning company, D-ID, it can generate a facial video from a single image.
By mapping a pre-recorded driver video containing several sequences of gestures to a still photo image, it can replicate facial expression and head movements.
No doubt this will soon garner a lot of meditation attention and elevate the debate on deep fakes and generative videos. Deep fakes have been a hot topic for quite some time now, but it’s interesting to see how commercial applications of this technology will evolve. Already there’s a lot of debate of the future of movies with deep fakes replacing real actors. It’s fairly apparent that even deceased actors can appear in new movies. New laws giving both living and dead actors rights to their digital personas are being enacted to counter this trend. At the EuroGraphics conference last year, Disney research presented a paper on face swapping and it has the potential to bring back dead actors in a realistic way. Their paper describes an extension to the GAN training process known as a progressive growing GANs that provides stable training of generator models that can output large high-quality images.
Interestingly the MyHeritage tool provider, D-ID, specializes in using deep learning to de-identify faces. Humans can still identify the face but by removing redundant facial biometrics its algorithm manages to do a pretty good job of fooling most facial recognition tools. Here’s a good explanation of how it works. Anonymizing facial recognition has a lot of interesting applications so expect a lot more development in this space.
Deepfakes and gans are just a few trending topics in the world of AI and data science. To learn more about the future of the industry, register now for ODSC East 2021 – tickets are 30% off right now.