OpenAI finally launched Sora, its artificial intelligence (AI) video generation model, on Monday. In February, the company previewed Sora to select individuals, and now, it released a different variant of the model dubbed Sora Turbo. Sora can generate videos in 1080p resolution which can be as long as 20 seconds. The AI model has been deployed on a standalone platform which is currently available as a website. Notably, Sora is currently only available to paid subscribers of ChatGPT with specified rate limits.

OpenAI’s Sora AI Video Generation Model

In a blog post, the AI firm announced the launch of Sora and detailed the capabilities of the model. Sora was first unveiled earlier this year, and the model has been repeatedly delayed. The company had stated that the reason behind the delay was strengthening the safety and privacy parameters of the model.

However, after a delay of nearly nine months, OpenAI has launched Sora as a standalone platform which can be accessed here. It is currently only available to ChatGPT Plus and Pro subscribers. Those without subscription cannot create a new account on the website currently. Meanwhile, Plus users are limited to 50 videos at 480p resolution or fewer videos at 720p every month.

ChatGPT Pro subscription, which was recently introduced at $200 (roughly Rs. 16,970) a month, will let users generate videos with “10x more usage, higher resolutions, and longer durations.” However, just like “fewer videos”, the company did not quantify what would entail under high resolutions and longer durations.

Sora can currently generate videos in widescreen, vertical, and square aspect ratios. Users can also upload their videos and images to extend, remix, and blend the content into generated videos. The AI model also allows generating videos from scratch using text prompts. Additionally, a storyboard interface lets users set particular inputs for each frame.

Coming to technicalities, OpenAI explained that Sora is a diffusion model, where the AI has the foresight of many frames at a time to keep the content consistent over the 20-second period. The AI model uses a transformer architecture, and takes recaptioning technique from DALL-E 3.

OpenAI also highlighted the details about the model data. The company claimed that it sourced a wide range of data from the public domain, via its data partnerships, and data from people working with the model. The public data was said to be collected from machine learning datasets and web crawls.

The company also partnered with Shutterstock Pond5 and commissioned datasets to generate proprietary data for the AI model. Finally, data for Sora was also collected from AI trainers, red teamers, and employees.

To minimise the risks associated with a realistic AI video generation model, OpenAI is adding both visible watermark as well as metadata as per the standards set by the Coalition for Content Provenance and Authenticity (C2PA). The company also claimed that it has added protections in the model for media uploads that include people.

The AI firm also stated that Sora will be blocked from generating videos containing damaging forms of abuse such as child sexual abuse and sexual deepfakes. Additionally, the number of uploads people can make will be limited at launch.

OpenAI’s Sora video generation model, which was unveiled in February but has not been released yet, appears to have been briefly leaked online. On Tuesday, an anonymous group hosted an artificial intelligence (AI) video model on Hugging Face and claimed it to be OpenAI’s Sora. The model details highlighted that the backend server was the AI firm’s domain, and the model name suggested it was the Turbo variant. The AI model was said to be listed on the platform for three hours before its access was removed.

OpenAI Sora AI Video Model Might Have Been Leaked

A Hugging Face listing appeared on November 26 and claimed to be publicly giving access to OpenAI’s Sora model. The anonymous group behind the move also created a front-end using which anyone could generate AI videos. While Gadgets 360 staff members were not able to generate videos using it, several social media users have posted videos generated using the tool.

The AI model could generate 10-second-long videos in 1080p resolution and the videos carried the distinct OpenAI watermark, which led to the belief that the group’s claim about it being Sora was likely true. Other evidence pointing towards it being Sora includes the backend server which was listed as “https://sora.openai.com/backend/video_gen” and the variant name mentioned as “Turbo”.

Explaining the reason behind leaking and posting the early version of Sora online, the group said that OpenAI used artists as “unpaid R&D” and made them provide the company with free bug tests, training data, validation tokens, and positive marketing.

“Hundreds of artists provide unpaid labour through bug testing, feedback and experimental work for the program[..]offering minimal compensation which pales in comparison to the substantial PR and marketing value OpenAI receives,” it said on the Hugging Face listing.

The group also said that while they were not against the use of AI as a tool for art, they did not agree with OpenAI’s artist programme. The group added that this was their way to urge the company to be more open, artist-friendly, and supportive of art. They also shared a petition that can be signed by those who agree with the message.

Notably, the group was part of the three hundred artists who received unlimited access to Sora to be its early testers, red teamers, and creative partners. For three hours, the front-end of the AI model was said to be working, then the group highlighted that the AI firm had shut down Sora’s early access for all artists.

A company spokesperson told TechCrunch that Sora remains in a research preview as OpenAI works to balance the AI model’s performance with robust safety measures. The spokesperson also added that participation as an early tester is voluntary “with no obligation to provide feedback or use the tool.” The artists were asked to use the AI model responsibly and refrain from sharing confidential details till the model’s launch, the statement concluded.