She is good at both mammoths and human eyes.
The startup OpenAI announced the creation of the Sora model, specializing in generating videos from text descriptions. The current version of Sora can generate videos up to one minute long.
Unlike similar neural networks of competitors, Sora produces videos of very high quality. OpenAI engineers ensured that each frame looked more or less natural, and that the frames themselves blended well with each other.
As OpenAI explains, the model understands not only what objects look like, but also how they interact with each other in the real world. Sora is capable of generating videos with multiple characters, following exactly the user’s instructions.
Judging by the demo videos, the neural network copes well with animations and complex materials like fur or snow. The creatures generated by Sora also look natural – even if they aren’t particularly realistic themselves.
At the same time, Sora has disadvantages: the neural network has difficulty simulating scenes with complex physics and does not always understand cause-and-effect relationships. For example, a neural network can draw a video where there is no bite mark left on the cookie.
At the current stage, teams assessing the product for potential risks will have access to Sora. In addition, designers, artists and directors will have access to the neural network: they will provide feedback to OpenAI.
OpenAI cannot yet say when Sora will be available to all users of the company’s services.
This is interesting