OpenAI's latest model takes text messages and turns them into "complex scenes with multiple characters, specific types of movement," and more.

OpenAI is launching a new video generation model, and it’s called Sora. The AI company says Sora “can create realistic and imaginative scenes from text instructions.” The text-to-video model allows users to create photorealistic videos up to a minute long — all based on prompts they’ve written.

Sora can create “complex scenes with multiple characters, specific types of movements, and precise subject and background details,” according to OpenAI’s introductory blog post. The company also notes that the model can understand how objects “exist in the physical world,” as well as “accurately interpret props and generate compelling characters that express vivid emotions.”

Advertisement

The model can also generate a video based on a still image, as well as fill in missing frames in an existing video or extend it. The Sora-generated demos included in OpenAI's blog posts include an aerial view of California during the gold rush, a video that looks like it was shot from inside a Tokyo train, and others. Many have some signs of AI — like a suspiciously moving floor in a video of a museum — and OpenAI says the model “can struggle to accurately simulate the physics of a complex scene,” but the results are generally pretty impressive.

A couple of years ago, text-to-image generators like Midjourney were at the forefront of models’ ability to turn words into images. But recently, video has started to improve at a remarkable pace: companies like Runway and Pika have shown off their own impressive text-to-video models, and Google’s Lumiere figures to be one of OpenAI’s main competitors in this area as well. Like Sora, Lumiere gives users text-to-videotool and also allows them to create videos from a still image.

Sora is currently only available to “red teamers” who assess the model for potential harm and risk. OpenAI is also offering access to select visual artists, designers, and filmmakers for feedback. It notes that the existing model may not accurately simulate the physics of a complex scene and may not correctly interpret some cases of cause and effect.

Earlier this month, OpenAI announced that it was adding watermarks to its text-to-image tool DALL-E 3, but notes that they “can be easily removed.” Like its other AI products, OpenAI will have to contend with the consequences of fake, AI-generated photorealistic videos being mistaken for the real thing.