【動画】高層階の大きな窓から見下ろす素晴らしい景色の部屋

OpenAI, which develops the GPT series of large-scale language models and the safety DALL-E of image generation AI, has announced AI “Sora,” which can generate up to a minute’s video from text.Sora says she can create a movie after “understanding how it exists in the physical world” with “complex scenes that contain accurate details of multiple characters, specific types of motion, subjects and backgrounds.”AI that generates images from input text (prompts) like Midjourney, Stable Diffusion, and DALL-E has appeared since around 2022, and as technology advances, AI that generates videos from prompts has also appeared.However, these video generation AIs often produce conflicting movies because they are difficult to accurately simulate physical conditions in complex situations and do not understand the cause and effect of movements.For example, if a boy produces a movie where he eats cookies, for some reason there is no bite mark left on the cookie that he was supposed to have bitten.In addition, there are cases where confusion occurs due to spatial information such as left, right, and depth.

Sora is a noise diffusion model using Transformer architecture similar to the GPT model, and physical simulation is performed to generate a movie.As a result, it is possible to create dynamic images such as rotating the camera or moving around the subject.”We believe Sora will serve as the foundation for models that can understand and simulate the real world, and will be an important milestone in achieving AGI (General Purpose Artificial Intelligence),” OpenaAI commented.

The following is part of the movie that was actually generated by Sora, and the string directly above the movie is the prompt used to generate it.