Sora (model)
OpenAI's text-to-video generation model
Sora is a text-to-video generation model developed by OpenAI that creates realistic, high-definition video clips from natural-language text prompts, using a diffusion transformer architecture trained to generate video as sequences of…
Definition
Sora is a text-to-video generation model developed by OpenAI that creates realistic, high-definition video clips from natural-language text prompts, using a diffusion transformer architecture trained to generate video as sequences of spacetime patches.
Overview
Sora generates video by extending diffusion-model techniques, originally developed for still-image generation, into the temporal dimension. Rather than treating video generation as a simple sequence of independently generated frames, Sora represents videos and images as collections of smaller units called spacetime patches, analogous to how text transformers process sequences of text tokens. A transformer-based diffusion model then learns to iteratively denoise these patches into coherent video conditioned on a text prompt, allowing the model to maintain consistency of subjects, backgrounds, and motion across a generated clip. At its announcement, Sora demonstrated the ability to generate clips up to roughly a minute long featuring complex scenes with multiple characters, specific types of motion, and detailed backgrounds, at a level of visual fidelity and temporal consistency that significantly exceeded prior text-to-video systems. OpenAI also highlighted Sora's ability to extend existing videos, fill in missing frames, and generate video from a still image, positioning it as a more general video-generation and editing tool rather than only a text-to-video generator. OpenAI released Sora gradually, first to a small group of safety testers and visual artists to study risks such as misinformation and misuse before wider availability. Its release intensified competition in the generative video space, with other labs and startups (such as Luma AI's Dream Machine and Stability AI's Stable Video Diffusion) racing to offer comparable capabilities, and it raised prominent discussions in the creative industries about the implications of AI-generated video for film, advertising, and visual effects work.
Key Features
- Text-to-video generation from natural-language prompts
- Diffusion transformer architecture operating on spacetime patches
- Generates clips with strong temporal and subject consistency
- Supports video extension, frame interpolation, and image-to-video generation
- Capable of complex multi-character scenes and detailed backgrounds
- Developed and released by OpenAI with staged rollout for safety testing
- Intensified competition among text-to-video model providers
- Raises significant questions around misinformation and creative-industry impact