Ever wondered what Mona Lisa would look like rapping? Microsoft launches VASA-1 AI bot that can make images talk - with eerily realistic results

Trending 1 month ago

The bound betwixt what's existent and what's not is becoming ever thinner acknowledgment to a caller AI instrumentality from Microsoft. 

Called VASA-1, nan technology transforms a still image of a person's look into an animated clip of them talking aliases singing. 

Lip movements are 'exquisitely synchronised' pinch audio to make it look for illustration nan taxable has travel to life, nan tech elephantine claims. 

In 1 example, Leonardo da Vinci's 16th period masterpiece 'The Mona Lisa' starts rapping crudely successful an American accent. 

However, Microsoft admits nan instrumentality could beryllium 'misused for impersonating humans' and is not releasing it to nan public. 

Microsoft's caller tool VASA-1 tin make clips of group talking from a still image and audio of personification talking - but nan tech elephantine isn't releasing it immoderate clip soon

VASA-1 takes a fixed image of a look – whether it's a photograph of a existent personification aliases an artwork aliases drafting of personification fictional. 

It past 'meticulously' matches this up pinch audio of reside 'from immoderate person' to make nan look travel to life. 

The AI was trained pinch a room of facial expressions, which moreover lets it animate nan still image moreover successful existent clip – truthful arsenic nan audio is being spoken. 

In a blog post, Microsoft researchers picture VASA arsenic a 'framework for generating lifelike talking faces of virtual characters'. 

'It paves nan measurement for real-time engagements pinch lifelike avatars that emulate quality conversational behaviors,' they say. 

'Our method is tin of not only producing precious lip-audio synchronisation, but besides capturing a ample spectrum of emotions and expressive facial nuances and earthy caput motions that lend to nan cognition of realism and liveliness.' 

In position of usage cases, nan squad thinks VASA-1 could alteration integer AI avatars to 'engage pinch america successful ways that are arsenic earthy and intuitive arsenic interactions pinch existent humans'. 

But experts person shared their concerns astir nan technology, which if released could make group look to opportunity things that they ne'er said. 

VASA-1 requires a fixed image of a look – whether it's a photograph of a existent personification aliases an artwork aliases drafting of personification imaginary. It 'meticulously' matches this up pinch audio of reside 'from immoderate person' to make nan look travel to life

Microsoft's squad said VASA-1 is 'not intended to create contented that is utilized to mislead aliases deceive'

Another imaginable consequence is fraud, arsenic group online could beryllium duped by a clone connection from nan image of personification they trust. 

Jake Moore, a information master astatine ESET, said 'seeing is astir decidedly not believing anymore'. 

'As this exertion improves, it is simply a title against clip to make judge everyone is afloat alert of what is tin and that they should deliberation doubly earlier they judge correspondence arsenic genuine,' he told MailOnline. 

Anticipating concerns that nan nationalist mightiness have, nan Microsoft experts said VASA-1 is 'not intended to create contented that is utilized to mislead aliases deceive'. 

'However, for illustration different related contented procreation techniques, it could still perchance beryllium misused for impersonating humans,' they add. 

'We are opposed to immoderate behaviour to create misleading aliases harmful contents of existent persons, and are willing successful applying our method for advancing forgery detection. 

'Currently, nan videos generated by this method still incorporate identifiable artifacts, and nan numerical study shows that there's still a spread to execute nan authenticity of existent videos.' 

Microsoft admits that existing techniques are still acold from 'achieving nan authenticity of earthy talking faces', but nan capacity of AI is increasing rapidly. 

Regardless of nan look successful nan image, nan instrumentality tin shape realistic facial expressions that lucifer nan sounds of nan words being spoken  

According to researchers astatine Australian National University, clone faces made by AI seem much realistic than quality faces. 

These experts warned that AI depictions of group thin to person a 'hyperrealism', pinch faces that are much in-proportion, and group correction this arsenic a motion of humanness.

Another study by experts astatine Lancaster University recovered clone AI faces look much trustworthy, which has implications for online privacy.  

Meanwhile, OpenAI, nan creator of nan celebrated ChatGPT bot, introduced its 'terrifying' text-to-video instrumentality Sora in February, which can make ultra-realistic AI video clips based solely connected short, descriptive matter prompts. 

This framework of an AI-generated video of Tokyo created by OpenAI's Sora shocked experts pinch its 'terrifying' realism  

In consequence to nan punctual 'a feline waking up its sleeping proprietor demanding breakfast', Sora returned this film 

A dedicated page connected OpenAI's website has a rich | assemblage of nan AI-made films, from a man stepping connected a treadmill to reflections successful nan windows of a moving train and a feline waking up its owner.

However, experts warned it could swipe retired full industries specified arsenic movie accumulation and lead to a emergence successful heavy clone videos starring up to nan US statesmanlike election. 

'The thought that an AI tin create a hyper-realistic video of, say, a leader doing thing untoward should ringing siren bells arsenic we participate into nan astir election-heavy twelvemonth successful quality history,' said Dr Andrew Rogoyski from nan University of Surrey.

A investigation insubstantial describing Microsoft's caller instrumentality has been published arsenic a pre-print. 

Four of these faces were produced wholly by AI... tin YOU show who's real? Nearly 40% of group sewage it incorrect successful caller study 

Recognizing nan quality betwixt a existent photograph and an AI-generated image is becoming progressively difficult arsenic nan deepfake tech gets much realistic.

Researchers astatine nan University of Waterloo successful Canada group retired to find whether group tin separate AI images from existent ones.

They asked 260 participants to explanation 10 images gathered by a Google hunt and 10 images generated by Stable Diffusion aliases DALL-E – 2 AI programs utilized to create deepfake images – arsenic existent aliases fake.

The researchers noted that they expected 85 percent of participants to beryllium capable to accurately place nan images, but only 61 percent of group guessed correctly.

Read more 


Copyright © PAPAREAD.COM 2024

TECHNOLOGY