OpenAI restricts the generation of videos that include real people or imitate existing artists with its video generation AI 'Sora'



OpenAI released 'Sora Turbo,' the production version of its video generation AI Sora, on December 9, 2024. However, OpenAI has made it clear that the feature that can generate videos that incorporate real people will not be made available to anyone other than a very small number of users until further testing is completed.

Sora System Card | OpenAI

https://openai.com/index/sora-system-card/

Why OpenAI is only letting some Sora users create videos of real people | TechCrunch
https://techcrunch.com/2024/12/09/openai-is-only-letting-some-sora-users-create-videos-of-real-people/

You can learn more about the officially released Sora Turbo by reading the article below.

OpenAI officially releases video generation AI 'Sora', available with ChatGPT Plus and Pro plans - GIGAZINE



One restricted feature is the ability to generate videos based on uploaded photos and footage of real people, and while OpenAI has given a small percentage of Sora's users access to the feature, it says it has no plans to roll it out more broadly until it can fine-tune its approach to safety.

OpenAI also states that 'generated videos containing real people will not be allowed to be shared to the Sora public feed.'



'Early feedback from artists shows that the ability to generate videos from photos and videos of real people is a powerful creative tool. But because of the potential for misuse, we're taking a phased approach and will not roll it out to all users for now,' OpenAI said in a blog post.

Will Peebles, technical staff member at OpenAI and research lead on Sora, said in the presentation, 'We have clear ambitious goals at OpenAI. We want to prevent illegal activity with Sora, but we also want to balance it with creative expression. We know this is an ongoing challenge. We're not going to be perfect on day one. We're going to start a little modestly, and we welcome your feedback if our moderation isn't perfect.'



Additionally, OpenAI is implementing a filter to check whether the generated videos contain people under the age of 18. If they do, OpenAI claims to apply 'stricter standards' for moderation of sexual, violent, and self-harm content.

In addition, to avoid copyright infringement claims, the app also has a mechanism to rewrite prompts when trying to generate videos in the style of a current artist. In addition, all videos generated by Sora contain C2PA -compliant metadata that indicates that they are AI-generated. However, Tech Crunch, an IT news site, commented that 'the metadata can be removed.'

OpenAI has only stated that 'Sora was trained using a combination of 'diverse datasets,'' which are said to include publicly available data, proprietary data accessed through partnerships with vendors, and custom sets developed in-house. In April 2024, OpenAI reportedly used more than 1 million hours of YouTube videos for training.

It turns out that OpenAI used over 1 million hours of YouTube videos to train AI models - GIGAZINE

in Software,   Web Service, Posted by log1i_yk