Introducing Sora 2

Sora 2 represents a significant step-function change in video generation, building on the foundation of Sora 1, which was considered the “GPT-1 moment for video generation.” The new model is presented not just as a tool, but as “the most powerful imagination engine ever built,” integrated into a new social application.

Core Improvements in the Sora 2 Model

Sora 2 introduces several fundamental improvements over its predecessor, marking a leap forward in realism and capability.

• Integrated Audio Generation: For the first time, Sora simultaneously generates video and synchronized audio. This is a general-purpose system capable of producing multi-speaker dialogue in various languages, sound effects, and entire soundscapes.
• Enhanced Physical Understanding: The model demonstrates a much more robust understanding of physical interactions, dynamics, and complex collisions. Examples cited include challenging scenarios like an Olympic gymnastics routine or a backflip on a wakeboard, which now appear extremely natural. The model’s grasp of motion, physics, and body mechanics is described as state-of-the-art.
• Improved Steerability and Coherence: Sora 2 is more adept at generating longer, coherent narratives with multiple shots in a single generation. This addresses a common limitation in prior systems which often required a shot-by-shot approach, making it easier to tell a complete story in one go.
• Wide Stylistic Range: The model can generate content across a diverse spectrum of aesthetics, from photorealism to anime and everything in between. This versatility prevents the output from collapsing into a single, homogenous style.

The “Cameo” Feature: Personalized Generation

A flagship feature unique to Sora 2 is “Cameo,” which allows users to insert themselves, their pets, or even objects into any generated scene.

• Mechanism: By observing a short clip of a subject, the model gains a deep understanding of it and can then inject it into any prompt “as if it were just another text token.” This capability emerges from the underlying world simulation models.
• User Control and Safety: Creating a Cameo requires a liveness check and recording a dynamic audio prompt to prevent impersonation. Users retain full control over their likeness with granular permissions:
  • Only I can use my cameo.
  • People I approve.
  • Mutuals.
  • Everyone.
• Ownership and Customization: Users are treated as owners of any content generated with their Cameo and have the right to delete it. They can also provide preferences (e.g., “always wears a gold chain,” “no skinny jeans”) to guide the model on how they should be portrayed.

The Sora App: A New Social Medium

The Sora app is designed as a new product surface to encapsulate the model’s capabilities, particularly the Cameo feature. It is positioned as a new form of communication.

• AI-Generated Social Feed: The app features a familiar social media interface with profiles and follows, but all content is AI-generated by users. The experience is described as a “new medium” for connecting with friends.
• Content and Interaction:
  • The feed heavily prioritizes content from friends and family (“connected content”).
  • A “remix” feature allows users to instantly create their own variations on existing posts, facilitating participation in trends and collaborative storytelling.
  • A beta feature will allow users to guide their feed based on their current mood (e.g., “relaxing,” “animals”).
• Safety and Responsible Design:
  • Provenance: All exported content is visibly watermarked and incorporates C2PA standards to ensure it is clearly identifiable as AI-generated.
  • Moderation: The platform uses reasoning models to prevent the creation of harmful, violent, or X-rated content, with an initially conservative approach to moderation.
  • User Well-being: To combat “doom-scrolling,” the app includes features like no infinite scroll for U18s and nudges for adults to encourage creation over passive consumption.

Rollout and Availability

• Sora App: The iOS app will begin a US and Canada-based, invite-only rollout. New users will receive four invites to bring friends onto the platform, emphasizing the social experience.
• Web App and API: The existing sora.com web application will be updated with the Sora 2 model and will soon feature a “storyboard” tool for more granular, shot-by-shot control. An API for Sora 2 will also be launched in the coming weeks for developers.

The overall vision is that by building AI systems that deeply understand the physical world, we move closer to AGI. Along the way, these powerful models can be used to bring creativity, joy, and new forms of connection to people.

• Tags:
• Generative AI
• Feature
• OpenAI
• Audio Generation
• Video Generation
• Sora
• Sora 2
• Cameo
• World Simulation