(Credit: Shutterstock / SomYuZu)
Microsoft Research Asia new experimental AI tool can create some pretty believable deepfakes.
VASA-1 can take a still image or drawing of a person, pair it with an audio file, and create a lifelike talking face of them almost instantaneously.
The tool creates things that look pretty realistic and can mimic things like a person’s facial expressions and head motions from the photo and move their lips in a way that makes it look like they are in fact the one talking or singing. The group posted some videos of VASA-1 in action that are pretty wild to watch.
Up close, a trained eye can possibly tell that the head motions are a bit robotic, but the results are pretty believable. Which is one reason researchers are not releasing an online demo, API, or product that uses VASA-1 it until they can be certain it "will be used responsibly and in accordance with proper regulations.”
“We are opposed to any behavior to create misleading or harmful contents of real persons, and are interested in applying our technique for advancing forgery detection,” the group said, adding that videos currently created with the tool don't have identifiable artifacts.
Ultimately, the researchers see the tool as something that could be used to provide companionship and therapeutic support for people who need it, or to provide a “person” that someone could talk to in situations where AI is being used.
Earlier this week, Microsoft took its WizardLM-2 AI model offline within a day of its launch because the developers failed to complete toxicity testing prior to its release.