Researchers at the University of Washington have developed a new video creation method utilizing A.I., allowing them to create videos showing politicians and public figures saying things that they never actually said.
Quartz Media reports that a team of scientists at the University of Washington has developed a new video creation method, using artificial intelligence to render videos of politicians and celebrities saying things that they did not in fact say. To demonstrate this, the group used audio of former President Barack Obama and created multiple videos of the ex-President saying the sentence in multiple settings and locations with different facial expressions, head movements, and facial quirks.
This effect is achieved by having AI neural networks analyze 17 hours of footage from President Obama’s weekly address, analyzing the mouth and facial movements of Obama and then generating different mouth shapes from the separately analyzed audio of Obama speaking. These mouth shape outlines were then textured to match Obama’s lips and overlaid onto footage of Obama speaking; the videos were then retimed so that the mouth shapes would match with the former President’s expression in the video.
A video of the new software in action can be seen below,
This is not the first time a similar effect has been achieved using technology, in June of 2016 researchers at Stanford University created a similar technology, while not as advanced as the software created by the University of Washington, Stanford’s attempts are still quite convincing as can be seen in the video below,
The new software developed at the University of Washington, which can generate extremely high-quality videos from audio alone, may have many useful real life applications such as generating videos from audio so that the hearing impaired can read lips, or improving the realism of motion capturing and facial movements in modern video games and movies. But many worry that this sort of technology will be used to replicate the facial patterns of politicians, potentially being used as propaganda to smear candidates.
However, currently the A.I. used to create these facial patterns needs hours of footage to analyze in order to create a convincing likeness, that is why researchers chose former President Obama’s weekly address footage as a test subject as it provided hours of high-quality footage to work with. Generating similar footage of other public figures would not be quite so easy given current software limitations.