OpenAI has finally opened its much-hyped Sora video generator to the public — and its work isn’t exactly blowing us away.
One particular weakpoint: Sora’s attempts to generate videos of gymnasts result in a horrorshow of whirling and morphing limbs that would be more at place among David Cronenberg‘s body horror films than the Olympics.
As venture capitalist Deedy Das quipped in a thread showing the bizarre creations that “gymnastics is still very much the Turing test for AI video.”
Das went on to post three more uncanny Sora-generated gymnastics videos, and there’s more where that came from.
Echoing the venture capitalist, the account associated with the AI for Humans podcast joked that “gymnastics are the funniest way to break [S]ora.”
Indeed, in a video made with the prompt “gymnast flips five times very fast then lands in a bucket of mustard,” the same limb confusion as the previous videos is yet again on display — but this time, a waterfall of what looks like yellow paint is attached to the “feet” of the artificial gymnast, making the whole thing look all the slimier.
That same account also had Sora make a video of a Raygun-esque breakdancer that had a similarly creepy limb confusion effect.
sora has its own interpretation of Raygun’s breakdance from the Paris Olympics
prompt: famous Australian woman breakdances at the Paris Olympics pic.twitter.com/IefygMDCpT
— AI For Humans Show (@AIForHumansShow) December 11, 2024
Even in videos with less movement, where Sora often does significantly better, it struggles with spelling words correctly — a basic imagine generator issue that’s emerged as a tell for AI images.
In one post, a self-described “politically homeless bitch” claims that the quality of Sora’s videos is “mind-bending” — without pointing out that in the last of the four videos of a woman in fascist military uniform, the mask she’s wearing misspells “obey” as “oeybey.”
In an even more egregious example, vlogger Marquees Brownlee said that when he got early access to Sora, it kept outputting “garbled” text despite the often-photorealistic quality of its videos.
“How can [Sora] create a photorealistic human but can’t spell basic words?” another user observed in a post screenshotting Brownlee’s video.
While it certainly has its impressive aspects, Sora by and large seems to be way less sophisticated than its boosters would have had you believe before it launched publicly — a common theme among AI hypebeasts.
More on Sora: OpenAI Concerned About Illegal Activity on Sora, Releases It Anyway
Source link
lol
