OpenAI Unveils Sora: Advanced Text-to-Video Model
S
SlashPage
OpenAI has unveiled Sora, a cutting-edge generative video model that transforms concise text descriptions into detailed, high-definition film clips lasting up to a minute. This development represents a notable leap in text-to-video generation, showcased through sample videos that demonstrate Sora's impressive ability to understand complex 3D interactions and effectively handle occlusion. Despite strict secrecy conditions during the preview, OpenAI has not released a technical report or a demonstration of Sora in action, and there is no immediate plan for public release. The company is currently sharing Sora with a select group of safety testers and creative professionals to gather feedback and address potential misuse concerns.
Built upon technology from DALL-E 3, OpenAI's flagship text-to-image model, Sora combines a diffusion model with a transformer, allowing it to process video data across both space and time. The transformer's capability to handle long sequences of data, similar to its application in language models like GPT-4, enables Sora to be trained on diverse video types in terms of resolution, duration, aspect ratio, and orientation. While the showcased videos highlight Sora's strengths, including high-definition output and effective occlusion handling, OpenAI acknowledges the need for further refinement, particularly in ensuring long-term coherence.
Video generated with Sora Ai
OpenAI is attentive to potential risks associated with generative video models, including misinformation and deepfake misuse. To address these concerns, Sora includes filters blocking requests for violent, sexual, or hateful content, and a fake-image detector developed for DALL-E 3 is adapted for use with Sora. Industry-standard metadata tags are embedded in Sora's output to indicate how the video was generated. Overall, Sora showcases significant advancements in generative video models, but OpenAI remains cautious about deployment, emphasizing the importance of gathering feedback and ensuring safety before any public release.
/calhigh-machine-learning
Subscribe