Sora 2 improves physical simulation and integrates synchronized audio generation

OpenAI has announced the launch of Sora 2, the second version of its video and audio generation model. The system represents an evolution from the original model released in February 2024, with improvements in the physical accuracy of simulations and new control capabilities.

The company highlights that Sora 2 can generate complex sequences such as Olympic gymnastics routines or acrobatic movements on paddleboards, maintaining coherence in aspects like buoyancy and object rigidity. Unlike previous versions that modified reality to fulfill text instructions, the new model respects physical laws with greater fidelity. In the example provided by OpenAI, if a basketball player misses a shot, the ball bounces off the backboard instead of teleporting to the hoop.

The model integrates audio generation, including soundscapes, dialogue, and sound effects. It also allows following detailed instructions spanning multiple shots while maintaining virtual environment coherence. Among its capabilities is the insertion of real elements into generated scenes, a feature OpenAI calls "cameos" that allows incorporating people, animals, or objects from real videos into system-created environments.

OpenAI has developed an iOS application called Sora where users can create content, remix other users' generations, and utilize the cameos feature. The latter requires an initial video and audio recording to verify identity and capture the user's likeness, after which they can be inserted into any generated scene. The application includes a recommendation system based on language models that can be configured through natural language instructions.

Regarding safety measures, the company has established limits on daily generations visible to teenagers and parental controls through ChatGPT. Users maintain full control over their digital image, being able to revoke permissions or remove videos that include them. OpenAI notes that the planned monetization model consists of offering additional paid generations when demand exceeds available computing capacity.

Sora 2 is initially available in the United States and Canada free of charge with usage limits, while ChatGPT Pro subscribers will be able to access an experimental version called Sora 2 Pro. The company plans to expand the service to other countries and launch an API for developers. The previous model, Sora 1 Turbo, will remain operational.

Key points

• Sora 2 improves physics simulation accuracy compared to previous video generation models.
• The system generates video and audio in synchronized form, including dialogue and sound effects.
• The "cameos" feature allows inserting real people or objects into generated scenes after an initial recording.
• OpenAI has launched a social iOS app focused on content creation and sharing.
• The recommendation system can be configured through natural language instructions.
• Users maintain control over their digital image and can revoke permissions at any time.
• Includes daily usage limits for teenagers and parental controls via ChatGPT.
• Initially available in the United States and Canada free of charge with usage limits.