OpenAI’s generative video model Sora has officially launched – but the AI company is already noting its limitations.
This past Monday OpenAI launched Sora, an AI video generation model capable of creating FullHD videos from text prompts as well as user-uploaded videos and photos. OpenAI notes that Sora still has some limitations, such as unrealistic physics, but says the launch will provide time to develop safeguards for the emerging technology.
At launch, Sora is rolling out to paid subscribers using ChatGPT Plus and Pro plans, although the video generator is a standalone platform found at Sora.com. While the launch includes most countries where ChatGPT is already available, Sora is notably not yet available in the UK, EU or Switzerland.
The launch, however, was met with a larger than expected demand with some users reportedly unable to get the AI to work; OpenAI CEO Sam Altman noted on X that it was due to “underestimated demand” and that the company was working to get everyone access as fast as possible.
The AI enables users to generate video from text, but also enables generations built from existing images and video. At launch, the AI is limited to building FullHD generations that are up to 20 seconds long.
The Sora software includes a handful of tools to use both while generating a clip and for editing the generation after the fact. Settings include style presets, aspect ratios and resolution. Sora also enables users to generate one or multiple variations at once. After generation, the software includes editing tools like looping and mixing with another video.
While OpenAI first teased Sora in February 2024, the AI company says that the launch is an early version of Sora that has known limits. “The version of Sora we are deploying has many limitations,” the company wrote in the announcement. “It often generates unrealistic physics and struggles with complex actions over long durations. Although Sora Turbo is much faster than the February preview, we’re still working to make the technology affordable for everyone.”
The company says it's launching the video generator now in order to allow time for developing safeguards as the technology improves. “We’re introducing our video generation technology now to give society time to explore its possibilities and co-develop norms and safeguards that ensure it’s used responsibly as the field advances,” the company wrote.
At launch, the company says the tool has built-in blocks preventing the software from developing “child sexual abuse materials and sexual deepfakes.” The company has also limited the ability to generate videos of people based on a real photograph to only a limited subset of users for early testing.
AI platforms have largely been setting their own set of ethical guidelines. OpenAI’s DALL•E, for example, refuses to generate images of real people such as politicians but X’s Grok has no such limitation. OpenAI explained that, during testing, external “red teamers” were tasked with exploring risks such as identifying potential workarounds for graphic content.
For example, OpenAI said that these red teamers found that asking for certain medical or fantasy genres degraded the safeguards against depicting violence, enabling the developers to add more protections in those areas.
OpenAI says that Sora is a diffusion model that builds on some of the techniques and training from the image generator DALL•E. The videos used to train the AI were sourced from publicly available data, data partnerships such as those with Shutterstock Pond5, and feedback from employees and trainers. The company notes that the “publicly available data” includes machine learning datasets as well as videos sourced from web crawling.
The company also noted that it worked with “hundreds of visual artists, designers and filmmakers” for feedback on the early versions of the video generation model. OpenAI said that, after artist feedback, they removed the visible watermark on downloads. Videos are still embedded with C2PA or Coalition for Content Provenance and Authenticity tags, which notes in the file’s metadata that the video was generated by AI.
Last month, early glimpses of Sora leaked in what was touted as an “artist protest” from some of the tool’s early testers.
Accessing Sora requires an OpenAI Plus subscription, which is $20 / £16 / AU$31 a month and includes limited generation credits that amount to around 50 five-second videos. The Pro plan, introduced earlier this month, ups that to 500 20-second videos but costs $200 / £157 / AU$313 each month.
You may also like
For more on AI, read our guide on how to tell if an image is AI generated or how to prevent AI from using your images.
