OpenAI has unveiled its next-generation video generation model, Sora 2, and its invitation-based iOS app, Sora. Sora 2 claims to better follow the laws of physics and synchronize dialogue and sound effects with the screen for a more realistic experience. Users can create, remix, and share directly from the app. It also offers a cameo feature, allowing users to insert their own image and voice into scenes within their permission range after registering their portrait and voice. It's initially available on iOS in the US and Canada, with regional expansion planned for later. Web access will also be available once invitations open. Pro users will have access to the high-quality Sora 2 Pro model.
OpenAI explained that feed recommendations can be adjusted based on user language input. They prioritized exposure based on follower relationships and interactions, and emphasized a design that increases creation rather than time spent. They also disclosed their feed operating principles. They announced that protections would be in place for minor users, including viewing restrictions and enhanced cameo privileges.
Sora 2 reportedly reduces errors like object deformation and teleportation experienced in previous generations, and even reproduces failed scenes more realistically. The model simulates actions like a basketball bouncing off the backboard after a missed shot. Controls have also been improved, maintaining the world state of the scene while following multiple shot commands. Example footage includes challenging scenes like gymnastics, surfboard backflips, and figure jumps.
Sora reportedly features a TikTok-like social structure and is invitation-based. Users can join themselves and their friends in scenes as cameos and share them on their in-app feed. OpenAI acknowledged the risk of misuse of fake videos and said it implemented verification, permission revocation, and moderation. Initially, downloads will be available in the United States and Canada, with Android support planned for later.
The enthusiasm appears to be spreading quickly. There have even been reports of invitation codes appearing on secondhand trading platforms. OpenAI has warned against unauthorized sales.
This release further intensifies the generative video competition. Comparisons with similar features from Google's Veo 3 and Meta, as well as Runway, are inevitable. It remains to be seen whether Sora 2 can prove its superiority in physics simulation, voice synchronization, and controllability.
