Technical perspective: Photorealistic facial digitization and manipulation

ear appearance deformations of the mouth, in which plausible textures are retrieved instead of being rendered using a parametric model. Such an approach is particularly effective in producing a photorealistic output, as it bypasses the traditional and more complex rendering pipeline. While some limitations remain, such as the inability to control the head pose in the target video sequence, very convincing photorealistic facial reenactments are demonstrated on footages of celebrities and politicians obtained from YouTube. While the original intent of performance-driven video was to advance immersive communication, teleconferencing, and visual effects, the ease and speed with which believable manipulations can be created with such technology has garnered widespread media attention, and raised concerns about the authenticity and ethical aspects of artificially generated videos. Recent progress in artificial intelligence, such as deep generative models, is further accelerating these capabilities and making them even easier for ordinary people to use. For instance, Pinscreen’s photorealistic avatar creation technology requires only a single input picture and can be used to create compelling video game characters at scale, but face replacement technologies, such as DeepFake, have been exploited to create inappropriate and misleading video content. I highly recommend the following paper, as it is one of the first that promotes awareness of modern technology’s capability to manipulate videos, at a time in which social media is susceptible to the spread of doctored videos and fake news.