OpenAI has released Sora 2, the latest version of its video and audio generation model. Building on the original Sora from 2024, Sora 2 improves world simulation and can generate complex scenes with realistic physics. Unlike earlier models, it follows real-world dynamics, such as objects bouncing correctly, and can maintain consistent world states across multiple shots. It works in cinematic, realistic, and anime styles, and produces sound effects, speech, and background audio alongside the visuals.
The model can also incorporate real-world elements. Users can insert a person, animal, or object from a video into a Sora-generated environment, with accurate appearance and voice. While still imperfect, Sora 2 shows progress in scaling neural networks for video simulation.
Sora 2 powers a new social iOS app, Sora, where users can create and remix videos, and add themselves or friends into scenes using a “cameos” feature. The app is designed with safety controls, including limits for teens, moderation for inappropriate content, and full user control over cameo videos.
Sora 2 is initially available in the U.S. and Canada through the app and sora.com, with plans to expand. An experimental higher-quality version will be accessible to ChatGPT Pro users.
Leave a comment