Runway is developing real-time video generation technology designed to eliminate waiting periods between prompts and video output. The system aims to stream video content as users provide descriptions, minimizing the time required to generate the first frame.
The technology relies on GWM-1, a General World Model introduced in December 2025 that builds upon the Gen-4.5 architecture. This model processes camera movements, audio, and robot commands, while a related system, Solaris, generates user interfaces in response to voice or click inputs.
To address the issue of compounding visual errors, Runway trains its models on their own outputs to learn self-correction. This approach mirrors techniques used by Decart for its MirageLSD model. The company notes that real-time generation shifts compute loads from training to inference.
Runway envisions applications in education, gaming, and robotics, including synthetic data generation for training autonomous vehicles. A research preview developed with Nvidia and demonstrated at the GTC conference runs on the Vera Rubin platform, achieving a first-frame latency of under 100 milliseconds.
