OpenAI
- OpenAI has launched a new tool, Sora, that generates AI videos based on user text.
- It's currently in beta with safety experts and a limited group of creators.
- But the first — admittedly hand-picked — results look eerily realistic.
AI-created images are about to get a lot more real-looking.
On Thursday, OpenAI showed off a new tool called Sora that can generate strikingly realistic videos based on users' prompts.
The videos created by Sora can be up to a minute long, and can consist of "complex scenes with multiple characters, specific types of motion, and accurate details," the company announced.
In addition to text, Sora can also generate videos from still images and "extend" existing videos — including the ability to "fill in missing frames," according to OpenAI.
The product is initially rolling out to red teamers (experts in fields like misinformation, hate, and bias, who work with OpenAI to improve product safety) as well as "a number of visual artists, designers, and filmmakers to gain feedback," according to the company.
(Star YouTuber MrBeast seemingly joked about being concerned by the tech, and asked whether he was going to lose his job)
OpenAI shared several impressive examples of Sora's capabilities, including a video of a woman walking in the neon glow of Tokyo streetlights, wooly mammoths galloping through the snow, an aerial view of a church on the Amalfi Coast, and a cartoon monster curiously kneeling before a melting candle.
That said, the company also acknowledged the nascent tool's weaknesses, including faultiness when "simulating the physics of a complex scene" and not understanding "cause and effect" — like, for instance, a cookie that appears whole after someone's taken a bite.
The model also mixes up left and right and struggles with descriptions of events that take place over time, OpenAI said.
Before rolling out widely, OpenAI also said it's drilling down on safety by building tools to detect videos that Sora generated, and rejecting prompts that request violence, sex, hateful imagery, celebrity likeness, and third-party-owned IP.
"We'll be engaging policymakers, educators, and artists around the world to understand their concerns and to identify positive use cases for this new technology," the company said.