OpenAI has released Sora 2, its new and improved video generation model. The standout feature is “cameos,” where users can place a realistic avatar of themselves, or any human, animal, or object, in a generated clip.
To generate a cameo, a user just needs to capture a short, one-time video and audio recording of themselves. This can then be saved and brought into different videos through OpenAI’s new standalone “Sora” app, which is being launched alongside Sora 2.
The app is designed to be a new social networking platform, powered by Sora 2, rather than a simple AI video generator. Rumours of OpenAI joining the social app market were swirling back in April.
As well as creating their own videos, friends on the Sora app will be able to edit each other’s and bring in each other’s cameos. “At a time when all major platforms are moving away from the social graph, we think cameos will reinforce community,” the Sora team wrote in their announcement.
At the same time, OpenAI is conscious of preventing nonconsensual use of a person’s likeness, which has become a real problem in the AI industry that lawmakers have not gotten a handle on. Users can control who has access to their cameo, revoke permissions at any time, remove any videos featuring it, and review all videos that include their cameo. The social platform will also be invite-only, which could help to put a lid on bot and scam accounts.
The new Sora app will provide each user with a personalised TikTok- or Instagram Reels-style feed for discovering new videos. According to OpenAI, this feed will be shaped by the people users follow, the content they interact with, their ChatGPT history, and what the algorithm predicts will inspire them to create their own videos, rather than simply keeping them watching.
If users aren’t satisfied with their feed, they can adjust it using natural language prompts, and the Sora app will also periodically poll users on their well-being and suggest adjusting their feed. “Concerns about doomscrolling, addiction, isolation, and RL-sloptimized feeds are top of mind,” the team wrote.
They added that “a lot of problems with other apps stem from the monetization model incentivizing decisions that are at odds with user wellbeing,” alluding to how tech companies are inclined to design their products to maintain user attention for as long as possible so they can extract more ad revenue.
While OpenAI is averse to this, it will eventually offer users the option to pay to generate an additional video if demand exceeds available compute resources. Paid ChatGPT Pro users will also be among the first to access its experimental, higher-quality Sora 2 Pro model.
Sora 2 can generate even more realistic videos
OpenAI states that Sora 2 is “more physically accurate, realistic, and more controllable” than its predecessor. It can accurately model the dynamics of buoyancy and rigidity when generating, say, a gymnast doing a backflip on a paddleboard. It will also follow the laws of physics, even if that means disobeying its prompt.
“For example, if a basketball player misses a shot, the ball may spontaneously teleport to the hoop,” the Sora team said. “In Sora 2, if a basketball player misses a shot, it will rebound off the backboard.”
OpenAI says that Sora 2 is better at following instructions over multiple shots while maintaining the world state. This has also been included in the promotion for several AI video models this year, including Runway’s Gen-4, Google’s Genie 3, Mirage, and Manus.
OpenAI is taking teen safety seriously
Protecting children from AI is currently top of mind after a string of tragic incidents involving chatbots. OpenAI itself is being sued by the parents of a 16-year-old boy who died by suicide, who claim ChatGPT reinforced his darker thoughts and offered information about how he could end his life.
This week, the company added a host of new parental controls to ChatGPT, including a review system that may alert parents if their child expresses thoughts of self-harm while using the chatbot. There are also similar parental controls for Sora, which allow parents to turn off algorithm personalisation and infinite scroll, and choose whether their teen can send and receive direct messages. This does assume, however, that a parent is aware of their existence and how to use them.
In addition, OpenAI is placing limits on the number of AI videos teens can watch per day and implementing stricter permissions for cameos, as well as employing teams of human moderators to review any instances of potential bullying.
The Sora iOS app is now available for download, and accounts will be granted access on a rolling basis. Currently, only users based in the US and Canada can sign up, but access will be expanded to additional countries soon. Sora 1 Turbo remains available, and any videos will remain in a user’s Sora library.
Creators will need to actively opt out of having their copyrighted character recreated by Sora 2, and OpenAI is notifying talent agencies and studios of this.
The post OpenAI’s Sora 2 Lets You Put Yourself in an AI-Generated Video appeared first on eWEEK.
No Responses