OpenAI introduced a new tool on Thursday that can generate videos based on text prompts.
The tool, named Sora, draws inspiration from the Japanese word for “sky” and is capable of producing realistic videos up to a minute long, following user instructions on content and style. According to a post on OpenAI’s blog, Sora can generate videos from a single image or extend existing footage by adding new elements.
The blogpost explains, “We are training AI to understand and reproduce real-world dynamics, with the goal of creating models that can assist users in addressing real-world challenges.”
An example provided by OpenAI showcases a video created from the prompt: “A movie trailer featuring the adventures of a 30-year-old astronaut wearing a red wool knitted motorcycle helmet, set against a blue sky and salt desert background, with a cinematic style, shot on 35mm film, and vibrant colors.”
Access to Sora has been granted to a limited group of researchers and video creators for thorough testing to ensure compliance with OpenAI’s terms of service, which prohibit content containing “extreme violence, sexual material, hateful imagery, celebrity likenesses, or others’ intellectual property.” The tool is currently accessible to researchers, visual artists, and filmmakers. CEO Sam Altman has shared video clips on Twitter in response to user prompts, featuring videos reportedly generated by Sora, each bearing an AI authorship watermark.
Previously, OpenAI introduced the still image generator Dall-E in 2021 and the generative AI chatbot ChatGPT in November 2022, which quickly gained a user base of 100 million. While other AI companies have released video generation tools, their models often produce short clips that differ significantly from the provided prompts. Google and Meta have confirmed ongoing work on generative video tools but have not made them publicly available yet. Google recently announced an experiment to enhance ChatGPT’s memory capacity for retaining more user interactions.
OpenAI has not revealed the exact footage used to train Sora or the sources of the training videos, except for mentioning to the New York Times that the dataset includes videos from publicly available platforms and copyright-licensed sources. The company has faced legal challenges alleging copyright infringement in training its generative AI models, which learn from vast amounts of web-scraped data to replicate the content present in those datasets.
