OpenAI has officially introduced Sora, its new artificial intelligence video generation model, on Monday. Following a preview in February for select users, the company released an enhanced variant known as Sora Turbo. This model is capable of producing videos at a 1080p resolution, with durations extending up to 20 seconds. Sora operates on a dedicated platform, currently accessible through a website, but is exclusively available to paid subscribers of ChatGPT, complete with specific usage limits.
Launch of OpenAI’s Sora AI Video Generation Model
In a blog post, OpenAI announced the launch of Sora, elaborating on its features. The model was first revealed earlier this year, but its release faced multiple delays. The company indicated that this postponement was due to efforts to strengthen the model’s safety and privacy standards.
After a nearly nine-month delay, OpenAI has made Sora available as a standalone platform found here. Access is currently limited to ChatGPT Plus and Pro subscribers, and new users without a subscription cannot create accounts on the website at this time. Plus subscribers can create a maximum of 50 videos at a resolution of 480p or fewer videos at 720p each month.
The newly introduced ChatGPT Pro subscription, priced at $200 (approximately Rs. 16,970) per month, offers users the ability to generate videos with significantly greater allowances for usage, along with higher resolutions and longer video lengths. However, specifics regarding what constitutes “higher resolutions” and “longer durations” remain unspecified.
Sora enables users to generate videos in various formats, including widescreen, vertical, and square aspect ratios. Additionally, users can upload their own videos and images, allowing for content remixing, extending, and blending with the AI-generated visuals. The model also supports the creation of videos from scratch based on text prompts, complemented by a storyboard interface where inputs for each frame can be set.
On a technical note, OpenAI detailed that Sora functions as a diffusion model, allowing the AI to have insight into multiple frames simultaneously, ensuring consistent content throughout the 20-second videos. The model employs transformer architecture, drawing on techniques from DALL-E 3 for enhanced performance.
OpenAI provided insights into the data sourcing for the model, claiming it gathered a diverse array of information from public domains, data partnerships, and contributions from individuals collaborating on the project. The public data acquisition involved machine learning datasets and web crawls.
Moreover, the company partnered with Shutterstock, Pond5, and created commissioned datasets to generate proprietary data for Sora. Data was also compiled from AI trainers, specialized evaluators, and OpenAI staff members.
To mitigate risks linked with the realistic generation of AI videos, OpenAI is implementing visible watermarks and metadata in accordance with standards set by the Coalition for Content Provenance and Authenticity (C2PA). Additionally, the company stated that Sora will be restricted from producing videos that depict harmful forms of abuse, such as child sexual abuse and sexual deepfakes. At launch, there will also be limits placed on the number of uploads for users.
