Runway Aims to Shift AI Video to Controlled Live Mode

Runway is working on generating continuous video streams driven by voice, gestures, or commands. The company aims to deliver the first frame in under 100 ms on an Nvidia platform, without announcing a timeline. Behind this promise lie issues of GPU costs, interactive applications, and a major challenge: preventing small visual errors from turning into distortions.

Aiming for the First Frame Under 100 ms, but No Timeline Announced

In March, Runway presented a research preview at the GTC conference regarding a real-time model developed with Nvidia. This system operates on the Vera Rubin platform and is designed to deliver the first frame in less than 100 milliseconds. No availability timeline has been provided by Runway. According to the company, transitioning to real-time shifts the computational load towards usage: each frame must be produced quickly enough to keep up with playback, on hardware shared across multiple sessions.

Robotics, Education, and Autonomous Driving Rely on Interactivity

Runway presents interactive applications as the major long-term use case for generated video. Education, gaming, and robotics require responses as fast as the user's attention. Evaluating robots or autonomous vehicles also necessitates environments that react instantly to edge cases, and Runway has already introduced GWM Robotics to generate synthetic training data. In the same vein, Waymo relies on Waymo World Model, based on Genie 3 and adapted for traffic, to simulate events that its fleet has never encountered, such as a tornado, an elephant, or a flooded neighborhood. According to Waymo, its system virtually travels billions of miles before facing these scenarios on the road.

Advanced GPU Cost Savings to Broaden Use Cases

Runway highlights a potential reduction in costs with real-time generation. Faster models would consume less GPU time, improving their profitability. According to the company, the cost per output at a given quality determines the viability of an application, and instantaneous generation would lower this threshold, making previously unprofitable uses feasible. Runway also sees immediate iteration as a way to bridge the gap between an idea and its execution, focusing the user on creative direction rather than waiting.

Limiting the Snowball Effect of Visual Errors

Unlike text, video stacks its outputs: small errors can lead to distortions if not corrected. Runway describes this phenomenon as a central problem of approaches inspired by LLMs and claims to train its models on their own outputs to learn to correct discrepancies rather than amplify them. The company's real-time models use all previous frames as context, which heightens the need for consistent robustness at each image. Others are following similar approaches: Decart has subjected MirageLSD to images with defects during training, and Google DeepMind indicates that Genie 3 can maintain coherent interactive universes over several minutes at 24 frames per second in 720p.

What Runway Has Already Unveiled: GWM-1, Characters, and Solaris

Runway began this shift in March with Runway Characters. The system is based on GWM-1, presented as the company's first General World Model, introduced in December 2025, derived from Gen-4.5, capable of generating frame-by-frame videos and accounting for camera movements, robot commands, or audio. A few weeks ago, the company also showcased Solaris, which produces frame-by-frame user interfaces with responses to clicks and voice. More broadly, Runway has shared a preview of research aimed at videos streamed as they are described, rather than obtained after a wait. Currently, the models still impose invite-wait-render cycles, with restarts in case of dissatisfaction. The company claims that its users primarily waste time generating and revising, and aims for a minimal time to the first frame before continuous streaming.