OpenAI previews Sora for generating video from text

OpenAI has previewed Sora, an AI model that generates videos from written instructions. Announced on 15 February 2024, the research system can produce clips up to a minute long, according to the company.

An early research preview

Access is initially going to safety testers and selected visual artists, designers and filmmakers. OpenAI is seeking feedback from creative professionals before a wider release. This is a research preview, rather than a public video-making service.

The company says it is testing the model with outside experts and developing detection tools. It plans to add C2PA provenance metadata if Sora is deployed in an OpenAI product.

One model, multiple video formats

The technical report describes a diffusion transformer trained on images and videos together. It compresses visual information into smaller patches, then learns to generate content by removing noise. That representation accommodates different clip lengths, resolutions and aspect ratios.

OpenAI demonstrates landscape and portrait output, alongside the ability to animate a still image. The same model can extend existing video forwards or backwards, create loops and connect separate clips through generated transitions.

The researchers use detailed training captions and expand short prompts into fuller descriptions, connecting language instructions to the visual model.

These are capabilities shown in OpenAI’s research materials, not results from a Sulat test. For filmmakers and designers, they outline several possible creative workflows beyond producing a single clip from a written prompt.