Microsoft shows AI that generates hyper-realistic deepfakes from 1 photo

O VASA-1 is the latest AI from ecosystem, designed to create hyper-realistic deepfakes from a single photo or drawing of a person. The tool impresses with its ability to accurately reproduce facial expressions and head movements, giving the illusion that the person in the photo is actually expressing something. In demonstrations presented by ecosystem, VASA-1 impressed by replicating lip movement synchronized with the audio and creating expressions not present in the original photos, resulting in extremely convincing videos.

The new tool raises concerns about the ethical use of deepfake technologies, as its ability to create realistic videos with little input information could potentially be used to create fake news, defamatory videos and even hoaxes. Therefore, the ecosystem is working to ensure that the VASA-1 be used responsibly and ethically.

What is VASA-1 and how does it work?

It seems true, but the video above is just a deepfake. O VASA-1 represents a significant advance in generating realistic talking faces through artificial intelligence. O VASA-1 is capable of producing videos of talking faces from a single still image, be it a photo or drawing, of an individual and an audio clip of speech. These resulting videos feature not only lip movements synchronized with the audio, but also a wide range of natural facial dynamics and head movements, achieving a high level of realism and vividness.

Unlike previous methods, the VASA-1 approaches the generation of talking faces holistically, considering all possible facial movements, such as expressions, eye movements, and blinks, as unique latent variables. Furthermore, the VASA-1 incorporates optional conditioning cues, such as primary gaze direction, head distance, and emotional offset, into the learning process. This makes generative modeling of complex distributions more controllable and increases generation accounting. In the video below, for example, the VASA-1 demonstrated different gaze direction options in the deepfake.

One of the most impressive aspects of VASA-1 is its ability to generate talking faces in real time, which makes it ideal for interactive communication applications. By balancing video generation quality with computational efficiency, VASA-1 It significantly surpasses existing methods, bringing us closer to a future where AI-powered digital avatars can interact with us as naturally and intuitively as interactions with real humans.

The representation of VASA-1 can change the appearance, three-dimensional position of the head and facial dynamics, which allows separate control of attributes and editing of generated content. This means that, even with a single input photo, it is possible to generate videos of talking faces with different movement sequences or apply different photos to the same movement sequence, resulting in a wide variety of customization possibilities and control over the generated content. .

Dangers of deepfakes

O VASA-1, despite its possible positive applications, also presents significant risks related to the creation of deepfakes. This technology can be misused to create extremely convincing fake videos in which a person is depicted doing or saying something that never occurred. These deepfakes have the potential to cause serious harm, such as spreading misinformation, manipulating public opinions, defaming individuals, and even inciting social or political conflicts.

Especially during election periods, the technology could generate problems arising from malicious use with creations of political deepfakes. With this technology, it is possible to create videos of politicians or public figures making false speeches or carrying out compromising actions. These videos can be used to influence elections, undermine public trust in leaders and institutions, and generate political instability.

The use of Deepfakes has aroused the attention and concern of several governments around the world. In Brazil, the TSE banned the use of deepfakes in elections, with the measure being approved in February this year. This prohibition aims to prevent the manipulation of information and protect the integrity of the electoral process, preventing false videos and audios from being used to harm or favor candidates. The improper use of deepfakes may result in the revocation of the mandate or registration of candidacy.

China, in particular, has pioneered comprehensive regulation of the use of these technologies. Its legislation, broader than that adopted by some Western governments, is seen as an instrument to maintain social stability. It explicitly prohibits the creation of deepfakes without consent and requires clear identification of AI-generated content.

An alarming example occurred after the Russian invasion of Ukraine, when a deepfake video was widely circulated on social media. In it, Ukrainian President Volodymyr Zelensky appeared to order his troops to surrender, something that never actually happened. Furthermore, deepfakes can be used more widely in everyday situations, such as creating fake videos of celebrities, friends or family, creating confusion and damaging the reputation of the people involved.

Release forecast

A ecosystem recognizes the risks associated with VASA-1 and is committed to ensuring that the tool is developed and used responsibly. Due to the potential ethical, privacy and security issues that VASA-1 can generate the ecosystem There is no release date for the general public yet.

The company is actively working to implement security and control measures that help mitigate the risks of technology misuse. Developers are working to improve the authenticity of generated videos and develop deepfake detection methods that can help combat misuse of the technology, before considering its release to the general public.

See this and other news on Showmetech TRIO:

See also:

Sources: PCMag, ecosystem e shorts

reviewed by Glaucon Vital in 22 / 4 / 24.