Artists briefly leak access to OpenAI’s Sora video generator – SiliconANGLE

A group of artists has leaked access to Sora, an OpenAI artificial intelligence model designed for video generation that is currently in private alpha.

The artists made Sora’s application programming interface accessible via Hugging Face, Quartz reported today. OpenAI blocked access to the API after about three hours.

Sora debuted in February and is currently not available to the public. It allows users to generate videos up to one minute in length with natural language prompts. A prompt can contain several sentences describing what objects a clip should depict, how those objects should interact and other details.

Under the hood, Sora is based on the same transformer neural network architecture that underpins OpenAI’s language models. It uses a technique called diffusion to generate clips. Diffusion produces video content through a multistep process: Sora first creates an early version of a clip that contains a large amount of noise and then gradually polishes the video in several increments. 

When OpenAI debuted the model in February, it detailed plans to share it with a limited number of artists through an early access program. The company stated that the goal is to collect feedback on how Sora can be made more useful for creative professionals. OpenAI also shared the model’s API with a number of red teamers, cybersecurity experts who focus on identifying vulnerabilities and other issues in AI models.

Sora’s API was posted to Hugging Face by a group of about 20 artists who participated in the early access program. They explained that they leaked access to the model because they found fault in how OpenAI managed the program. “What we don’t agree with is how this artist program has been rolled out and how the tool is shaping up ahead of a possible public release,” they wrote.

The group took issue with the fact that Sora-generated videos must be approved by OpenAI before they can be shared. Additionally, the artists criticized an initiative through which the ChatGPT developer plans to screen films created by some early Sora testers. The initiative offers “minimal compensation which pales in comparison to the substantial PR and marketing value OpenAI receives,” the artists wrote on Hugging Face.

The company said in a statement that “hundreds of artists in our alpha have shaped Sora’s development, helping prioritize new features and safeguards. Participation is voluntary, with no obligation to provide feedback or use the tool.”

OpenAI has not yet provided a release date for Sora. It did, however, share plans to add C2PA support in the event that it decides to make the model commercially available. C2PA is a technology that makes it easier to determine if a video was generated by AI.

If the company decides to commercialize Sora, it may also release new versions that address some of the limitations in the model’s current iteration. It divulged in February that Sora sometimes struggles to “simulate the physics of a complex scene.” On occasions, the model also misinterprets prompts that include “spatial details” such as the direction in which an object should move. 

OpenAI’s DALL-E series of image generation models, which is already commercially available, has received several major upgrades since its release. The company has also built it into ChatGPT. It’s possible OpenAI will add a similar integration to Sora if and when it decides to make the video generator commercially available. 

Image: OpenAI