Sora has been announced by OpenAI, the San Francisco-based company famously known as the creator of Chat GPT and Dall-e image generator. Sora is the latest upgrade in generic artificial intelligence that can make short realistic videos from written prompts, which was unveiled on Thursday.
What can Sora do?
It can create high-quality videos of up to a minute in length from a command of short text (prompt). In addition, Sora can extend a short video or convert a photo into a video. Examples include a range from animated monsters and kangaroos to realistic videos such as a woman walking down a street in Tokyo or a cinematic video of a spaceman on a salt desert.
The tool is currently in a testing phase for a limited number of creators and is not yet publicly available.Â
What differentiates Sora?
Meta, Google, startup Runaway ML, and other companies have already demonstrated the text-to-video technology before Sora. However, Sora demonstrates a high-quality video, and it raises ethical and societal concerns for some groups of people.Â
OpenAI addressed this problem by quoting,
“We are working with red teamers — domain experts in areas like misinformation, hateful content, and bias — who will be adversarially testing the model,” the company said in its statement. “We’re also building tools to help detect misleading content, such as a detection classifier that can tell when Sora generated a video.”
Open AI also added they are creating new techniques to prioritize ethics and ensure the implementation of safety precautions that are used in programs such as DALL·E 3 for Sora. For example, Open AI says their text classifier will check and reject text input prompts violating their usage policies. The usage policies include extreme violence, hateful imagery, sexual content, celebrity likeness or the IP of others. Open AI has developed an image classifier that reviews the frames of every video generated to ensure that it adheres to policies before it is provided to the user.Â
Safety Tool to Test Sora
Open AI addressed that the current Sora model has weaknesses, including confusing right and left or failing to maintain visual continuity throughout the video.
Open AI said, “We’ll be engaging policymakers, educators and artists worldwide to understand their concerns and to identify positive use cases for this new technology.”
The company said they would implement adversarial testing called red-teaming on Sora to strengthen safety. Red-teaming will make dedicated users try to make the platform fail by producing inappropriate content.
Image Sources 1
