• TWEAK
• Software

OpenAI Sora impresses with wild AI Videos

Translate from : OpenAI Sora imponerer med vilde AI Videoer

OpenAI unveiled a tool Thursday that can generate videos from text prompts.

The new model, named Sora after the Japanese word for "heaven", can produce realistic footage up to a minute long that follows a user's instructions in terms of both subject and style. According to a blog post from the company, the model is also capable of creating a video based on a still image or augmenting existing footage with new material. Sora is able to generate videos up to 60 seconds long from text instructions with the ability to deliver scenes with multiple characters, specific types of movement and detailed background details.

"We teach AI to understand and simulate the physical world in motion with the goal of training models that help humans solve problems that require interaction with the real world," the blog post says.

OpenAI said it intends to train the AI models so they can "help people solve problems that require interaction with the real world."

A video included among several initial samples from the company was based on the call: "A movie trailer featuring the adventures of the 30-year-old spaceman wearing a red woolen knitted motorcycle helmet, blue sky, salt desert, cinematic style, shot on 35mm film, vivid colors."

The company announced that it had opened access to Sora to a handful of researchers and video producers. The experts would review the product, testing it for vulnerabilities to circumvent OpenAI's terms of service, which prohibit "extreme violence, sexual content, hateful images, celebrity likeness or other people's IP," according to the company's blog post. The company only allows limited access to researchers, visual artists and filmmakers, though CEO Sam Altman responded to users' calls on Twitter after the announcement with video clips he said were made by Sora. The videos carry a watermark to show that they are made by AI.

At the same time, OpenAI said Sora is still a work in progress with clear "weaknesses", especially when it comes to spatial details of a prompt that mixes left and right as well as cause and effect. It gave the example of creating a video of someone taking a bite of a cookie, but not having a bite mark right after.

The company launched image generator Dall-E in 2021 and generative AI chatbot ChatGPT in November 2022 that quickly accumulated 100 million users. Other AI companies have launched video generation tools, although these models have only been able to produce a few seconds of footage that often has little relevance to their calls. Google and Meta have said they are developing generative video tools, although they have not released them to the public. On Wednesday, it announced an experiment to add deeper memory to ChatGPT so it could remember more of its users' conversations.

OpenAI did not disclose how much footage was used to train Sora or where the training videos might have come from, other than to say that the collection of training material contained videos that were both publicly available and licensed from copyright owners. The company has been sued several times for alleged copyright infringement in the training of its generative AI tools.