OpenAI Introduces Sora 2: Video Generator Gains Realistic Physics and Sound
The company announces a qualitative leap in video generation, comparing the new model to the transition from GPT-1 to GPT-3.5.
OpenAI has announced the first major update to its video-generation neural network, Sora 2. According to the developers, this is not just an iterative improvement but a significant evolutionary step that takes the technology to a new level.
Key innovations in Sora 2 affect several aspects:
- Realistic Physics Simulation. The neural network has learned to model object interactions more accurately, making the generated videos more plausible.
- Integrated Soundtracks. Videos are now created with background noise and even voiceovers, freeing users from the need to edit sound separately.
- Handling Complex Scenarios. The model has become better at understanding detailed instructions describing sequential scene changes.
- Personalized Deepfake Function. Users can upload a short video of their face and a voice sample, allowing Sora 2 to create content with their digital avatar. The technology is also applicable to animals and objects.
Alongside the model update, OpenAI launched a separate mobile application, Sora, which serves as a platform for creating, publishing, and collaborating on AI videos.
Access to Sora 2 will be rolled out gradually. The basic version will be free with “generous limits,” while ChatGPT Pro subscribers will receive unlimited access and priority use of the more advanced Sora 2 Pro model.