Disney presents high-resolution neural facial exchange research

Gif: Disney Research Studios (Youtube)

In a few years, the neural networkautomatic facial exchanges have gone from being slightly convincing to eerily believable. But through new Disney research, the neural face …the exchange is about to become a legitimate and highQuality tool for visual effects studios working on Hollywood blockbusters.

One of the biggest challenges of creatingFake videos, as they are known, are creating a vast database of a person’s facial images, thousands of different expressions and poses, which can be swapped into a target video. The larger the database and the higher the quality of the images, the better the face swap. But images (which are often headshots of famous people) are generally obtained from sources with limited resolution. Even a 4K video file can produce low-resolution facial images since small faces often appear in the general frame of a shot.

So the first step in generating really compelling deep fake videos is to get started With a high quality font. meA new document presented at the Eurographics 2020 Symposium on Rendering (yet another event held online this year), “High resolution neural face swap for visual effects Researchers at ETH Zurich and Disney Research Studios detail several new innovations and approaches to the automatic faceexchanges that produce megapixel results with sufficient quality and resolution to be used in actual film production.

The new algorithm created by the researchers begins by actually modifying the source (or the target video) to facilitate the exchange of alternate faces. Movement in the source footage is subtly stabilized and smoothed to eliminate potential problems, such as trembling lip that could potentially bypass the automated swapping process in a later step. The researchers also improved several other steps along the way, including combining the new face into the original through improved composition techniques to better match overall contrast. The algorithm even does a much better job of generating the necessary intermediate frames to create smooth results so that the new face doesn’t appear to jump when playing altered material.

Every day there seems to be a new use for machine learning that promises to streamline and speed up a task that has generally taken a long time to complete, and from the first deepFake videos began to hit the internet, visual effects artists have seen the potential of the work they do. Exchanges of faces are not uncommon in the film and television industry; oftenSometimes a double stunt will momentarily stare at the camera, requiring extensive post-production to ensure, even for a brief moment, that the person in the the screen looks exactly the way it should be.

Fixing these issues can often require fresh shots, or a combination of smart computer graphics and composition, which is never cheap. With this new research, existing images from the same session could be used to train the algorithm, which would then solve these problems on its own. But while overworked visual effects artists and budget-minded Hollywood producers might celebrate the new tool, it will also make it much more difficult to spotfake videos in nature. New approaches in this research will soon find their way to existing machines.learning tools at which point we can expect a new wave of deepPretend to flood the Internetand Now there is a good chance that we do not know that they are false.

G / O Media may receive a commission