Runway has published research into real-time video generation built on the GWM-1 model
Runway has presented research into a system that would stream video continuously from a text description instead of making users wait for a finished clip. It is built on the GWM-1 model and addresses the accumulation of errors by training on its own outputs; no availability date has been announced.
Runway has published research into a system for real-time video generation that would allow users to watch video being generated continuously from a text description, instead of the current process of entering a prompt and waiting seconds to minutes for a finished clip. According to Runway, the aim is to minimize the time to the first frame and then stream the video as the user continues to describe and modify it.
The system is built on the GWM-1 ("General World Model") model, which Runway introduced in December 2025 and which is based on the Gen-4.5 model. GWM-1 generates video frame by frame and accepts camera movements, commands to robots or audio as control inputs. According to Runway, a key problem with video models is that, unlike text models, they cannot go back and correct a frame that has already been generated — small errors therefore accumulate over time into significant distortions. The company addresses this problem by training the model on its own outputs, which is intended to teach it to correct its deviations instead of amplifying them.
Decart used a similar principle of training on erroneous outputs for its MirageLSD model. Google DeepMind states that its Genie 3 model maintains the consistency of an interactive world for several minutes at 24 frames per second at 720p resolution. According to Runway, real-time generation shifts the computational load from training to operation, because the model must create each frame fast enough to keep up with playback, on hardware shared by multiple sessions at once.
Runway sees the greatest potential in interactive applications — education, games and robotics — and mentions the GWM Robotics variant for generating synthetic training data for robots. Waymo uses a similar approach in its own Waymo World Model, built on the Genie 3 model and adapted for road traffic, to simulate scenarios that its vehicles have not yet encountered in real-world traffic. In March, Runway also presented a research demo of a real-time model developed jointly with Nvidia, running on the Vera Rubin platform, with the aim of delivering the first frame in under 100 milliseconds. Runway has not announced an availability date.
Why it matters
This is currently a research demo with no announced availability date, so there is no direct impact on everyday users or businesses yet. However, the direction of the entire field is relevant: according to Runway, faster generation reduces computing requirements and makes possible applications that would not be economically viable at current prices, and Waymo already uses the same principle of "world models" in practice to simulate traffic for training autonomous vehicles.
Relevant practical impact
What this means
For a business
According to Runway, faster models reduce GPU requirements and shift the threshold at which video applications become economically viable; Waymo already uses a similar approach to simulate scenarios for training autonomous vehicles, and Runway uses it to create synthetic training data for robots.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.