Researchers from ByteDance, the parent company of TikTok, have unveiled a groundbreaking AI system named OmniHuman-1, capable of producing some of the most lifelike deepfake videos seen to date.
The market is flooded with deepfake AI applications, offering a plethora of tools that can insert individuals into images or manipulate video to make it appear someone said something they didn’t. However, many deepfakes, particularly video variants, often stumble at the uncanny valley, revealing unmistakable clues that highlight AI’s involvement.
OmniHuman-1 appears to transcend these limitations, at least based on the selective examples shared by the ByteDance research team.
For instance, consider this fictional performance by Taylor Swift:
Here’s a non-existent TED Talk:
Additionally, here is a deepfake lecture featuring Einstein:
ByteDance researchers state that OmniHuman-1 requires only one reference image and audio, such as voice or music, to create a video of any desired length. The system allows users to modify the video’s aspect ratio and the subject’s “body proportion,” enabling adjustments on how much of the individual is visible in the generated footage.
Trained on an extensive dataset of 19,000 hours of video from undisclosed sources, OmniHuman-1 exhibits the capability to edit existing videos, including altering the movement of a person’s limbs. The results are often astonishingly convincing.
However, OmniHuman-1 is not without its challenges. According to ByteDance, using “low-quality” reference images can lead to subpar video results, and the system faces difficulties with certain poses. For example, observe the unusual gestures involving a wine glass in this video:
Nonetheless, OmniHuman-1 greatly surpasses prior deepfake technologies and hints at future advancements in the field. While ByteDance has yet to distribute the system, it is likely that the AI community will soon reverse-engineer such models.
The potential implications of this technology raise significant concerns.
Last year, politically motivated deepfakes proliferated across the globe. On election day in Taiwan, a group affiliated with the Chinese Communist Party released AI-generated audio misleadingly portraying a politician endorsing a pro-China candidate. In Moldova, deepfake videos falsely showed the country’s president, Maia Sandu, resigning. Meanwhile, in South Africa, a deepfake
Deepfakes are increasingly being weaponized for financial fraud. Consumers have been misled by deepfakes of celebritiescompanies have been defraudedDeloitte indicate that AI-generated content was responsible for over $12 billion in fraud losses in 2023, with projections estimating losses could reach $40 billion in the U.S. by 2027.
In February, a large portion of the AI community signed an open letter advocating for strict regulations surrounding deepfakes. With no federal law prohibiting deepfakes in the U.S., over ten states have passed laws against AI-facilitated impersonation. California is working on a law that would empower judges to mandate the removal of deepfakes and impose financial penalties on violators.
Detecting deepfakes remains a formidable challenge. Though some social media platforms and search engines have attempted to mitigate their dissemination, the proliferation of deepfake content continues to escalate at an alarming pace.
A survey conducted in May 2024 by the ID verification firm Jumio revealed that 60% of respondents encountered a deepfake within the last year. Seventy-two percent expressed constant anxiety about being deceived by deepfakes, while a significant majority supported legislation aimed at countering the spread of AI-generated fake content.
Stay informed with TechCrunch’s AI-focused newsletter! Subscribe here to receive it every Wednesday.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
