Runway has published research that turns video generation into something closer to a livestream. A person talks, and the model draws each frame as the description unfolds.
Today’s systems run in stages: prompt in, a pause of seconds or minutes, a file out, and a retry if it misses. Runway contends that this loop is where the hours vanish, and that collapsing the wait to a single first frame turns the job from reviewing output into steering it.
The engine behind the idea is GWM-1, which Runway calls its first General World Model and unveiled in December 2025. Gen-4.5 is the system it extends. Audio, robot commands and camera movement all work as controls, and the picture comes back one frame at a time.
Generating a frame at a time carries a hazard of its own. Each frame is conditioned on the one before it, so small deviations snowball into visible drift. Runway’s fix is to train the model on its own output rather than error-free data alone, so it learns to pull deviations back instead of magnifying them. Decart took a comparable route with MirageLSD. Genie 3, Google DeepMind’s world model, is credited with several minutes of consistent interaction. It runs at 24 frames per second in 720p.
Nvidia is a partner here. At the chipmaker’s GTC conference in March, the two showed a joint research preview aimed at the Vera Rubin platform, with a target of a first frame inside 100 milliseconds. Nothing has been scheduled for release.
The economics run both ways. Faster generation burns less GPU time per clip, lowering the quality bar at which an application pays off. It also shifts load from training to serving, since every frame must arrive in time for playback while hardware is shared across sessions.
