Here is the next generation of Openai video generators.
The company's latest video and audio generation model, SORA 2, upgrades its photorealistic video capabilities and is billed as a social platform for users to share, remix and discover AI-generated videos.
Introduced in livestream on Tuesday afternoon, SORA 2 represents the latest in the wave of multimedia generation AI tools that allow users to create increasingly realistic images, videos and audio.
Based on previous AI image generation technology, Sora 2 allows users to create “cameos” or guest appearances of themselves or others in video. This feature requires users to create one-time video and audio recordings to verify their identity.
Like many existing social media platforms, the SORA app introduces an algorithmic feed that displays videos tailored to the user's interests, based on topics that users may be involved with who they interact with. According to details published by OpenAI, the feed includes a “steerable ranking” system that allows users to further personalize what they want to see.
SORA 2 is based on the original SORA model released in February 2024.
Sora's first version sometimes struggles to represent realistic movements like basketball bouncing off the backboard, but the Open said on Tuesday that Sora 2 “it's better to follow the laws of physics.”
Announced on Openai's website, Demos featured photorealistic action shots of stunts, including gymnastics tumbling and skateboarding tricks, but the martial artist clips performed on Koi Pond featured staff who appear to be unable to retain their shape.
SORA 2 can also generate speeches that were not possible in previous versions of SORA themselves.
“This model is far from perfect and makes many mistakes, but it is a test that further expanding the neural network with video data will bring you closer to real-world simulations,” Openai wrote in its presentation.
On Tuesday, Openai research scientist Gabriel Petersson introduced the model's cameo capabilities by posting a video to X, jumping out of a cargo ship on a dragon, and running through the Openai office with a portrait of CEO Sam Altman.
The SORA app with SORA 2 can now be downloaded on iOS Systems, but only invites access to the service. Users can request access through the app.
Openai has shown that access could be slowly rolled out across the US and Canada, initially giving users “generous restrictions” to video creation. AI systems like SORA 2 require intense computing power, so businesses often place restrictions on user access to allow others to access the service.
“Until transparent, our only current plan is to provide the option to generate additional videos to ultimately pay the users some amount if there is too much demand for the available calculations.”
Openai also acknowledged concerns about the potential risks of the app and released another safety blog post in conjunction with a larger announcement.
All SORA generated videos contain industry standard metadata to clarify that they were made with AI, the company writes. Additionally, the user's portrait can only be converted to a cameo with consent and can be revoked at any time.
SORA teenage accounts are subject to parental control and time limits imposed on using the app.
“At creation, Guardrails try to block unsafe content before it is created by checking both prompts and output across multiple video frames and audio transcripts, including sexual material, terrorist propaganda, and self-harm promotion,” Openai wrote in the post.
To prevent potential risks, Openai said it is expanding its team of human moderators to review bullying content.
Openai has assembled the SORA 2 as a key step towards a more powerful AI system.
“The video model is getting really good, very quickly,” Openai said in a release statement. “General-purpose world simulators and robotic agents fundamentally reconstruct society and accelerate the arc of human progress.”
