Keanu Reeves neutralizes the robber: how enthusiasts shot a parody video about the actor using neural networks

It seems that Deepfake brought to a new level.

Authors from the YouTube channel Corridor Digital released a video with Keanu Reeves, stopping an armed robbery. According to the video plot, fans meet an actor in an American store where a robber finds himself.

Almost to the very end, the video looks quite realistic, but in fact it was made using the Deepfake technology. Keanu’s face was put on a similar stuntman, having trained the algorithm based on neural networks and processed the result for convincingness.

The Corridor Digital team spoke about the nuances of creating a video in a separate video . As the authors explained, they are Reeves’ longtime fans and wanted to beat his fame as the “best of people.”

The team had an idea to make a video about how the actor prevents robbery. Since they could not call Kean himself, they decided to use neural networks.

Before that, the authors already had experience with Deepfake when working on a video about Tom Cruise, but the video with Reeves turned out to be much more complicated. For example, since the actor has long hair, just using a wig was not enough, so they put a cap on the stuntman – so the problem was solved.

For the correct operation of the algorithm, it is also necessary that the face be lit as evenly as possible without sharp shadows. Therefore, the authors used on the set of additional lighting and avoided dark frames in moments with Keanu.

A separate problem was shooting in profile. Deepfake allows you to generate frames with a profile, but if a person on the source video is very different from the actor, then the result is a contour.

Therefore, the team needed a person who could simultaneously repeat the movements of Keanu Reeves and his voice, and also looked like him outwardly. In the end, one of the members of the Corridor Digital found on Facebook a stuntman, Ruben Langdon, who also works as a voice actor.

To generate a fairly realistic model of Keanu Reeves for the video, the authors trained the neural network two weeks in a row on various data with the participation of the actor. Mostly used his role in recent films.

To increase the quality of the face, the original video was driven through 2 thousand overlay iterations. The team could have made the result even clearer, but it would have taken more time.

Shooting a video took one day. During the filming, the authors did not even see a preliminary result, so they did a lot of duplicates in the hope that their quality would be suitable. The video was really shot on an ordinary smartphone – this is not artificial styling. Corridor Digital expects that Keanu will watch the video and hope that he will not sue them.

Back to top button