OpenAI has launched a groundbreaking text-to-video AI service: “Sora.”
Almost impossible to tell it's not real
The video content that is created just by describing in a few sentences what is desired is of such high quality that it is hard to imagine that it will not change the entertainment industry forever, and more. The biggest limitation is the length of the videos which are limited to one minute, so long, but then again this is beta.
For now, Sora is in a test phase where creators get access first: "Today, Sora becomes available to "red teamers" to be able to assess critical areas for potential damage or risk. We also provide access to a range of visual artists, designers and filmmakers to get feedback on how to advance the model to be most useful to creative professionals.”
OpenAI says it's already sharing its work with third parties to get feedback, meaning the technology will rapidly improve from what's already perceived to be real, at least from the clips shared by the company—however, mistakes can happen.
OpenAI:
“Sora builds on previous research with DALL·E and GPT models and uses the repetition technique from DALL·E 3, which involves generating highly descriptive captions for the visual training data. As a result, the model is able to follow the user's text instructions in the generated video more faithfully. In addition to being able to generate a video solely from text instructions, the model is capable of taking an existing still image and generating video from it, as well as animating the image's content with accuracy and attention to small details. The model can also take an existing video and expand it or fill in missing frames.”
Errors may occur
"Today's model has weaknesses. It can struggle to accurately simulate the physics of a complex scene and can struggle to understand specific cases of cause and effect. For example, a person may take a bite of a cookie, but afterward it may happen that the cookie is just as whole. The model can also be confused about details in space in a prompt, for example by mixing left and right, and can struggle with precise descriptions of events that take place over time, such as following a specific camera path," OpenAI reports in the press release about potential errors.
OpenAI specifies that they will make it possible to detect whether a video was created by Sora and will include C2PA metadata in the future if the model is included in an OpenAI product.