Produce video from a single photo with artificial intelligence
With the R&D studies carried out within the Samsung AI Lab, the technology of creating video from photos has been taken further by training with machine learning in a traditional CNN (Convolutional Neural Network) architecture, making it capable of producing video from even a single photo.
In 2017, an open source AI software called “Deepfakes” enabled any video to be simulated by placing someone else's head on someone else's body. Deepfakes were able to achieve results thanks to hundreds of videos and images using AI technology.
But what if we only have one photo when trying to create a digital clone? That's the future of new research being done in Samsung's AI lab. Starting with just a photo, Samsung's latest AI technique is able to animate a 2D image into a convincing full-motion video. Marilyn Monroe, Albert Einstein, and even the Mona Lisa herself were portrayed in this context:
Unlike Deepfakes, which uses a technique called GAN (Generative Adversarial Network), which gradually draws pictures with machine learning algorithms under AI, Samsung's system starts by scanning the landmarks on someone's face to understand how they are. Training the system is possible with a single photo, but as the number of photos increases, much more perfect results can be obtained.
References:
- Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, Victor Lempitsky1. "Few-Shot Adversarial Learning of Realistic Neural Talking Head Models", Samsung AI Center, Moscow & Skolkovo Institute of Science and Technology, 20 May 2019.
- https://stanford.edu/~shervine/l/tr/teaching/cs-230/cheatsheet-convolutional-neural-networks
