Hilarious AI Gymnastics Video Shows How Far the Technology Has to Go

A gymnast in red performing a routine on uneven parallel bars. She is suspended horizontally, with her legs extended out on one bar and her hands gripping the other. The background shows an indoor gymnasium with red seating.

A dystopian AI-generated video of gymnasts performing a routine proves how text-to-video technology still falls a long way short of being any good.

The humorous video has gone viral online and was created by Wright Bagwell using Luma AI’s free AI video generator Dream Machine.

The video shows the AI gymnasts moving in bizarre fashion; periodically appearing and disappearing while their bodies, often headless, split into a starfish of human limbs.

“If you look really closely at the background, you can tell it’s AI,” one joker quipped beneath Bagwell’s post.

Over on X (formerly Twitter), Tim Spalding speculated that part of the reason the AI footage is so bad is because the AI model wasn’t trained on enough gymnast videos. A hypothesis that is likely correct.

Bagwell also used Dream Machine to create a series of other AI-generated sports videos including cycling, wrestling, figure skating, table tennis, and more.

The terrible AI videos are reminiscent of the early AI images which were also famously janky such as those produced by OpenAI’s DALL-E 2.

Those early images that came out of DALL-E now have something of a cult fanbase with fans enjoying the imperfect nature of the AI content. Many were disappointed that OpenAI recently closed down DALL-E 2 after some had come to rely on its dreamy, shadowy aesthetic.

But much like DALL-E 2 — which has been greatly iterated on by Midjourney and DALL-E 3 — Luma AI’s Dream Machine will also be superseded by better models that are less whacky and more coherent.

AI video is still in its infancy. But it appears inevitable that this type of synthetic imagery will start competing with real videos and fooling people into believing they are real.

The implications of this are huge with fears that deepfake videos will be used by scammers to rob people or to convince people to vote in a certain way by having a politician act out something they didn’t do.

The most-hyped AI video generator is OpenAI’s Sora which was announced months ago. There is still no release date set as of writing. There is likely wrangling behind the scenes over fears of the technology and arguments about how Sora was trained which OpenAI has refused to reveal details about.