OpenAI has announced the release of Sora 2, its flagship video and audio generation model. The new model, which features enhanced physical accuracy, greater user control, and the ability to insert real-world elements into generated scenes, is being deployed through a new social application for iOS called “Sora.”
This release marks a significant step forward from the original Sora model launched in February 2024. OpenAI describes this advancement as a potential “GPT-3.5 moment for video,” indicating a substantial leap in capability and performance.
Key improvements in Sora 2
Sora 2 introduces several major advancements over its predecessor, moving closer to the goal of creating a functional world simulator.
- Enhanced physical accuracy: Previous video models were often “overoptimistic,” disregarding realistic physics to fulfill a user’s prompt. Sora 2 demonstrates a more grounded simulation of physical laws, accurately modeling outcomes like a missed basketball shot rebounding off the backboard rather than teleporting into the hoop.
- Advanced user controllability: The model can follow intricate, multi-shot instructions while maintaining the state of the generated world across different scenes and camera angles, allowing for more complex and coherent video narratives. It also shows proficiency across various aesthetic styles, including realistic, cinematic, and anime.
- Real-world element integration: Users can now inject elements from the real world into generated environments. By recording a video of a person, animal, or object, the model can place that element into any Sora-generated scene, accurately portraying its appearance and voice.
The Sora social app and Cameos feature
OpenAI is deploying the new model through a social iOS app designed for creating and sharing video content. The central feature of the app is “cameos,” which operationalizes the model’s ability to insert real-world elements.
To create a cameo, a user records a short video and audio clip within the app, which captures their likeness and voice for use in generations. Users have complete control over their personal likeness and can decide who is permitted to use their cameo. They can also revoke access or remove any video that includes their cameo at any time.
Focus on user wellbeing and safety
In launching the app, OpenAI has outlined measures to address concerns like digital addiction and social isolation.
- Feed philosophy: The app’s feed algorithm is designed to “maximize creation, not consumption,” prioritizing content from people the user follows and content likely to inspire their own creative work. OpenAI states it is “not optimizing for time spent in feed.”
- Teen safety: Specific safeguards for teenage users include default daily limits on the number of generations they can view and stricter permissions regarding the use of their cameos. OpenAI is also launching parental controls via ChatGPT to manage settings for teens’ accounts.
- Moderation: In addition to automated safety systems, the company is scaling up its teams of human moderators to review potential cases of bullying.
Availability and access
The Sora iOS app is now available for download in the United States and Canada, with plans to expand to other countries. Access is being rolled out on an invite-based system to encourage users to join with their friends.
- Pricing: The service will initially be free, with “generous limits to start.” OpenAI has stated that its only current monetization plan is to eventually allow users to pay for extra generations if demand exceeds available computing resources.
- Sora 2 Pro: Subscribers to ChatGPT Pro will have access to an experimental, higher-quality version of the model called Sora 2 Pro, which will be available on the sora.com website.
- API Access: OpenAI plans to release Sora 2 through its API for developers. The previous model, Sora 1 Turbo, will remain available.
OpenAI views the rapid improvement of video models as a crucial step toward developing general-purpose world simulators and robotic agents, presenting Sora 2 as “significant progress toward that goal.”