OpenAI unveils Sora, a model that creates videos up to one minute

OpenAI has unveiled Sora, an artificial intelligence system capable of creating videos up to one minute long from a written description. The announcement clearly raises the bar for video generators: the released samples show extended scenes, camera movements and multiple elements that maintain a notable degree of consistency.

For now, Sora is not available to the public. OpenAI has given it to external experts tasked with finding harmful uses and system flaws, as well as visual artists, filmmakers and designers who will provide feedback on its creative usefulness.

From text to moving scenes

Sora takes a natural-language prompt — for example, a scene, characters, a visual style and a camera movement — and produces a video sequence. It can also start with a still image and animate it, extend existing videos or generate frames to complete a sequence.

The company says the model can depict complex scenes with multiple characters, specific types of movement and environmental details. The published examples include animals, futuristic cities, natural landscapes and cinematic-style scenes. The ability to generate videos up to 60 seconds long matters because many earlier models focused on much shorter clips, where continuity errors are easier to hide.

The advance does not mean the system understands the world as a person does. OpenAI acknowledges significant limitations: Sora can make mistakes when simulating complex physical phenomena, confuse spatial relationships — such as left and right — or fail to correctly represent the cause and effect of an action. A person might appear to bite into a cookie, for example, without the cookie showing the corresponding bite mark.

A diffusion model with a transformer architecture

Technically, Sora combines two established ideas in generative AI. It is a diffusion model, meaning it learns to turn visual noise into a recognizable image or sequence step by step. It also uses a transformer architecture, the type of neural network that has driven language models such as GPT-4.

OpenAI divides videos into small units of visual data called patches, much as a text model processes fragments of a sentence. This approach allows it to work with videos of different lengths, resolutions and vertical or horizontal formats without having to build a separate system for each case.

The company also uses detailed descriptions of videos during training. This technique is intended to help the model follow user instructions more accurately and helps explain why Sora’s demonstrations respond to lengthy prompts about lighting, framing, characters or style.

The problem is not just technical

Improved audiovisual generation has obvious applications in advertising, film previsualization, education, video games and content creation. An agency could turn a script into a visual draft without organizing a shoot; a designer could test several ideas for a scene before producing it. But the same leap lowers the barrier to creating convincing fake videos.

OpenAI has begun testing with specialists in disinformation, hateful content and bias, among other risks. The company says it is developing tools to detect misleading material and that, if it deploys Sora in OpenAI products, it plans to add C2PA metadata. This technical standard makes it possible to attach information about a file’s origin and edits, although it cannot by itself prevent a video from being copied, cropped or shared outside platforms that read that data.

The restricted access reflects the fact that visual quality alone is not enough to launch a tool of this kind. Before it reaches users and businesses, OpenAI will have to show that its filters, provenance mechanisms and usage policies can address a problem that already affects generated images: quickly distinguishing synthetic content from a real recording when it circulates out of context.