OpenAI is showing off its new video generator, and it’s pretty impressive. CEO Sam Altman introduced the text-to-video model, which can make videos up to 60 seconds long. These videos feature detailed scenes, moving cameras, and multiple characters showing different emotions.
The first video Altman shared is of a couple walking down a snowy street in Tokyo, but it’s not clear what emotions they’re supposed to be showing. Still, the video looks very realistic and matches the detailed description given for it.
Previous attempts at AI video generation have had mixed results. Google’s “Lumiere” model, released last month, was an improvement but still had limitations.
However, the new model from OpenAI, called Sora, seems to be much better than Lumiere. Videos generated by Sora show various scenes in great detail, including wooly mammoths, a sci-fi movie trailer, a fluffy monster, and a paper-craft coral reef world.
While it’s unclear if the videos shared by CEO Altman were edited, those on OpenAI’s website were said to be directly generated by Sora without any changes.
There are still questions about how many videos OpenAI generated and how much computing power and time were needed to create them.
OpenAI acknowledges that Sora has some weaknesses. It might struggle with accurately simulating complex scenes and understanding cause and effect. For instance, it might not show a cookie with a bite mark after someone takes a bite.
Currently, Sora isn’t available to the public for another reason. OpenAI’s experts are still testing it to make sure it’s safe to use. They’re working on tools to detect misleading content generated by Sora, like a classifier that can spot videos made by the model.
It looks like the company wants to avoid repeating past mistakes and those of its competitors. They’re being cautious about releasing their model before making sure it won’t spread hateful lies.
“We’ll talk to policymakers, educators, and artists globally to understand their concerns and find good uses for this new technology,” the website says. “Even though we’ve done a lot of research and testing, we can’t predict all the good and bad ways people will use our technology.”
