OpenAI unveiled its groundbreaking text-to-video generator, Sora, last month, but it has yet to give Sora public availability. Currently, Sora is at the red teaming stage and is only available to select artists and creators who are testing the tool. For those unfamiliar, red teaming is the process where people test the tool to make sure it is safe, secure, and reliable. The goal is to identify vulnerabilities, biases, and other harmful issues.
In an interview with The Wall Street Journal (paywalled), Chief Technology Officer for OpenAI Mira Murati revealed that Sora will gain public availability “this year,” with a potential release within “a few months.” This tool, Sora, allows users to create hyperrealistic scenes from just text prompts.
Sora is mind-blowing, but not perfect yet
Sora-generated videos currently do not have audio in them. However, when asked about it Murati said that OpenAI has plans to integrate it somewhere down the line. It might not be Sora, but at least they have plans for the addition.
Just as many text-to-image generation tools, the videos from Sora will also be watermarked to avoid misuse. However, a watermark isn’t a perfect solution to the problem either. This is another reason the company has not yet released the model to the public, and reportedly won’t do it unless they consider it safe.
When asked about the CTO’s views on how it might affect jobs in the future, she stated the aim is not to replace people but to serve Sora as a tool to them that they can use to “edit and create with”. The editing of Sora-generated content can be significant, especially considering the inconsistencies in the videos. The Wall Street Journal showed some videos they have generated, pointing out some inconsistencies such as weird hands and fingers.
When questioned about the data used to train Sora, Murati remained vague, stating that it utilized publicly available or licensed data. While it remains uncertain whether videos from platforms like YouTube, Facebook, or Instagram were utilized, Murati confirmed that Sora incorporates content from Shutterstock, which is a partner of OpenAI.
OpenAI aims to make Sora available at a similar cost to its text-to-image model DALL-E
Murati highlighted one notable aspect, the increased cost of powering Sora compared to previous artificial intelligence models like DALL-E or ChatGPT. She notes that, unlike the text or image generation models, Sora is a research project at this point, which costs a lot more to operate. OpenAI will have to arrange the computing equipment it will require before serving it to the public.
Despite the higher expenses, OpenAI aims to make Sora available to the public at “similar” costs to DALL-E, emphasizing accessibility to its groundbreaking technology.
2024-03-14 15:09:23