On Tuesday, Google introduced Veo 3, a brand-new AI video synthesis design that can do something no significant AI video generator has actually had the ability to do previously: produce an integrated audio track. While from 2022 to 2024, we saw early actions in AI video generation, each video was quiet and generally really brief in period. Now you can hear voices, dialog, and sound impacts in eight-second high-definition video.

Quickly after the brand-new launch, individuals started asking the most apparent benchmarking concern: How excellent is Veo 3 at fabricating Oscar-winning star Will Smith at consuming spaghetti?

A short wrap-up. The spaghetti criteria in AI video traces its origins back to March 2023, when we initially covered an early example of dreadful AI-generated video utilizing an open source video synthesis design called ModelScope. The spaghetti example later on ended up being popular enough that Smith parodied it nearly a year later on in February 2024.

Here’s what the initial viral video appeared like:

Something individuals forget is that at the time, the Smith example wasn’t the very best AI video generator out there– a video synthesis design called Gen-2 from Runway had actually currently accomplished exceptional outcomes (though it was not yet openly available). The ModelScope outcome was amusing and strange sufficient to stick in individuals’s memories as an early bad example of video synthesis, helpful for future contrasts as AI designs advanced.

AI app designer Javi Lopez initially concerned the rescue for curious spaghetti fans previously today with Veo 3, carrying out the Smith test and publishing the outcomes on X. As you’ll observe listed below when you enjoy, the soundtrack has a curious quality: The synthetic Smith appears to be crunching on the spaghetti.

It’s a problem in Veo 3’s speculative capability to use sound impacts to video, most likely due to the fact that the training information utilized to develop Google’s AI designs included lots of examples of chewing mouths with crunching sound impacts. Generative AI designs are pattern-matching forecast devices, and they require to be revealed enough examples of different kinds of media to produce persuading brand-new outputs. If an idea is over-represented or under-represented in the training information, you’ll see uncommon generation outcomes, such as jabberwockies.