OpenAI Sora 2 System Card

OpenAI has released Sora 2, a state-of-the-art video and audio generation model designed to improve physical accuracy, visual realism, and audio synchronization. This model expands creative storytelling capabilities and serves as a research step toward simulating complex physical world dynamics.

Core Capabilities of Sora 2

Sora 2 introduces several technical advancements over the original Sora model, focusing on fidelity and user control. Key improvements include:

  • Physical Accuracy and Realism: The model achieves more accurate physics and sharper realism, allowing for generations that are better grounded in real-world dynamics.
  • Synchronized Audio: Sora 2 integrates audio generation that is synchronized with the video content.
  • Enhanced Steerability: The model follows user directions with higher fidelity, providing creators with more precise control over the output.
  • Expanded Stylistic Range: The model supports a broader variety of visual styles for imaginative and creative expression.

Deployment and Availability

Sora 2 is designed for accessibility across multiple platforms. Its availability includes:

  • Web Access: Available via sora.com.
  • Mobile Access: A standalone iOS Sora application.
  • API Access: Planned for future release.

Safety Framework and Risk Mitigation

OpenAI employs an iterative deployment strategy to manage the risks associated with high-fidelity video and audio generation, specifically targeting the nonconsensual use of likeness and the creation of misleading content.

Red Teaming and Iterative Safety

OpenAI collaborated with internal red teamers to identify emerging challenges and develop corresponding mitigations. The safety approach is iterative, prioritizing areas where context is critical or where risks are not yet fully understood.

Deployment Restrictions

To ensure a safe rollout, OpenAI has implemented the following safeguards:

  • Limited Access: Initial access is provided via limited invitations.
  • Upload Restrictions: The system restricts the use of video uploads and image uploads that feature photorealistic people.
  • Protection of Minors: Stringent safeguards and moderation thresholds are applied to all content involving minors.

Sources