OpenAI’s Sam Altman has announced the public release of Sora, the first-of-its-kind text and image-to-video generator. The model had been in beta testing since February.
It’s Day 3 of the ChatGPT-maker’s “12 days of OpenAI” new AI product announcements – and perhaps the most anticipated – with Monday’s launch of the latest version of its text-to-video generator “Sora Turbo.”
CEO and co-founder Sam Altman began the AI start-up’s third consecutive December live stream by showing off the latest Sora Turbo features in a demo accompanied by a gaggle of Sora product engineers.
The new Sora Turbo interface includes an Explore feed, a user Library, and a creative Storyboard tool, as well as a slew of new editing and creation features.
AI video will "change how we use computers"
Altman, who had been teasing how “so excited” he was about Sora Turbo’s official launch, said AI-generated video is important to the company, specifically highlighting three reasons.
“We love to make tools for creatives, it's important to our culture, it’s important to how we hope humans will use AI,” Altman cited as the first reason, noting a “new co-creative dynamic" witnessed emerging among early testers.
Secondly, Altman said the way AI understands how to use and generate video will ultimately change the way we use computers. “We don’t want the world to just be text,” he explained.
Finally, Altman said Sora itself “is critical to its AGI roadmap” referring to the developmental process of how “AI learns how to do things we need in the world.”
“We hope this early version of Sora will enable people everywhere to explore new forms of creativity, tell their stories, and push the boundaries of what’s possible with video storytelling,’ the AI start-up states.
By day's end, Sora will be available for free to all ChatGPT Plus and Pro subscribers in the US and most of the world, Altman announced.
ChatGPT Plus accounts will get “50 free generations per month,” at 480p resolution, or fewer videos at 720p, every month. Pro users will be afforded “500 fast generations (or fewer at high resolution) and unlimited in a slower generation mode” per month, said Altman.
details:
undefined Sam Altman (@sama) December 9, 2024
with an openai plus account, you get 50 generations per month.
with a pro account, you get 500 fast generations (or fewer at high resolution) and unlimited in a slower generation mode.
available in many countries, but not most of europe and the UK for now.
OpenAI said it was working on tailored pricing for different types of users for early next year. Unfortunately, ChatGPT users in the UK and most of Europe will have to wait until a later, unannounced date to partake.
Sora Turbo boasts new features
Sora Turbo is said to be a high-end accelerated and cheaper version of the Sora model initially released in February. Sora is a large-scale AI model that transforms a user's text prompt into 60 seconds of high-fidelity video. The service can generate videos from text prompts and pre-existing images or videos.
Capabilities include “generating video from text, animating images, and “a ton of video-to-video features, such as remixing into new styles, extending forwards and backwards in time, etc.,” said OpenAI lead Research Scientist Bill Peebles.
The visual introduction to Sora’s home feed – aptly named Explore – is a muti-box screen showing the random minute-long videos created by other Sora users to help provide inspiration. It features a Library homebase where users can view and organize their own video generations.
With the Storyboard feature, the idea is to take a video and edit it through text prompts to create another video at lightning speed, all by describing the environment and actions the user wants to occur on a timeline.
Users can also sequence the order, create endless loops, recut, blend videos together, and even allow the AI to generate new endings, beginnings, or middle sections for a video by simply leaving those areas blank on the Storyboard timeline.
The Sora crew demoed examples of a wooly mammoth trudging through a desert landscape turned into mammoth-robot creatures with a quick text directive as well as the creation of a 360-degree video of a lighthouse on an island surrounded by a moving ocean created from a simple image both in real-time.
Challenges remain
Sora team lead Aditya Ramesh said that the new model will allow users to push the boundaries of what is possible to do with video generations, but also reminded users that the early version of Sora will make mistakes.
For example, OpenAI lists the latest model “often generates unrealistic physics and struggles with complex actions over long durations.”
As for safety and copyright issues, OpenAI said it has taken initial steps to combat these types of problems.
According to the company, “particularly damaging forms of abuse, such as child sexual abuse materials and sexual deepfakes,” will automatically be blocked. Uploads of people will additionally be restricted during the launch period until more safeguards can be installed.
“All Sora-generated videos come with C2PA (opens in a new window) metadata, which will identify a video as coming from Sora to provide transparency, and can be used to verify origin,” and "watermarks are visible by default,” it said.
Last month, dozens of artists who signed an agreement with OpenAI to be early Sora beta testers had leaked an API version of the text-to-video AI generator on the AI developer website Hugging Face.
The leak was to protest being taken advantage of as ‘unpaid public relations puppets’ for the billion-dollar AI company.
Accompanied by an open letter, the artists accused OpenAI of luring them to be “early testers, red teamers, and creative partners” in an attempt to “artwash” the public, without proper compensation or credit for their work.
Curious folks can check out the “completely new experience” at OpenAI’s Sora.com.
Days 1 and 2 of OpenAI’s “12 days“ included safety improvements to its upgraded o1 System Card and the introduction of its new $200 per month ChatGPT Pro plan.
Your email address will not be published. Required fields are markedmarked