Original article was published by Louis (What’s AI) Bouchard on Artificial Intelligence on Medium
How DeepFakes are made
Like most recent AI-based applications, it uses deep neural network architectures to achieve this. As I said, it uses autoencoders, merged with GANs, which is especially good for computer vision applications like this one.
It consists of an encoder, which reduces an image to a lower-dimensional latent space and a decoder, which reconstructs the image from the latent representation.
The encoder is used to encode the person we want to imitate into the latent space. This latent space then contains the key features of their facial features and body posture to reproduce his facial and body movements.
Then, a model trained specifically for the target video is used to decode this latent space. This means that the target’s detailed information will be superimposed on the underlying facial and body features of the original video, represented in the latent space previously encoded.
GANs can be merged with such autoencoders to improve the results of this algorithm. It makes the deepfake constantly evolve and improve in realistically.