-
AI-generated video has come a long way.
-
In 2023, AI-generated videos went viral for depicting Will Smith eating spaghetti poorly.
-
OpenAI’s Sora and other models are coming close to passing the so-called spaghetti test.
It’s no longer your mom’s AI spaghetti.
In just two and a half years, AI video generation has progressed from struggling to depict Will Smith eating spaghetti to producing a lifelike video.
The unofficial benchmark test began in 2023 when a Reddit user posted a video of the Academy Award-winner eating spaghetti generated by ModelScope, a text-to-video AI model.
The initial results were horrifying. Will Smith looked nothing like his movie star self. Instead, he looked like a bad animation, complete with caricatured features that would be more at home on a tourist trap boardwalk. In some videos, he never actually consumed the spaghetti, failing to meet even the most basic premise of the test.
The failures highlighted the early limitations of AI-generated video and images, which sometimes produced people with 8 fingers or other anatomical imperfections.
Smith himself referenced the test in February 2024, posting a TikTok in which he ate spaghetti in almost as cartoonish a manner as the initial video.
A lot has changed since then, as SkyNews and others have recently pointed out.
In 2024, MiniMax, a Chinese AI model, made a much more accurate representation, but AI Smith’s chewing was still off. And at the very end of the clip, the noodles appear to levitate. In May, a user posted on X that he used Google’s Veo 3 to generate a new video. The problem with this one is that the noodles AI Smith is chewing sound way too crunchy. A later Veo 3.1-generated video looks even more realistic.
OpenAI’s Sora is widely regarded as the best AI video generator on the market. In fact, it is so good that soon after the launch of Sora 2 and its accompanying TikTok-esque mobile app in September, the company was forced to add more guardrails on third-party likeness and copyrights after a series of high-profile incidents involving SpongeBob and Martin Luther King Jr.
Google and Elon Musk’s xAI are racing to keep up. In July, xAI launched Grok Imagine, its text-to-video generator.
Passing the spaghetti test might be tougher now, as Hollywood and other rights holders intensify their efforts to prevent AI companies from infringing on their rights. Just days before Sora 2’s release, Disney, Universal, Warner Bros., and other rights holders filed suit in federal court against MiniMax.
Cameo, the personalized video company, sued OpenAI over its decision to name the core feature of the Sora app “cameos.” One of the reasons Sora can generate such high-quality videos, especially of non-public figures, is that users can upload facial scans to the app, which is the “cameo” feature. In November, a federal judge temporarily blocked OpenAI from using the word “cameo.”
Meanwhile, in Washington, some lawmakers are appalled that AI can now generate videos of them speaking words that they never once uttered.
Not everyone is shying away from AI video. Coca-Cola recently said it once again used AI to help generate its holiday ad, this time drawing on Sora, Veo 3, and Luma AI.
Read the original article on Business Insider
