Glonce

Runway demos real-time AI video streaming with GWM-1

researchThe Glonce editorsThe Decoder🔥71Glonce rating71Impact72Novelty65Relevance85Surprise58Players62Credibility75

Runway has shared research into real-time video generation, letting users stream video as they describe it instead of entering a prompt and waiting for a finished clip. Today's video models work in separate steps: enter a prompt, wait a few seconds or minutes, get a result, and start over if it isn't right. Runway says users repeatedly report losing the most time generating and revising videos, so it wants to minimize the time to the first frame, then stream video as users prompt it.

The work builds on GWM-1, Runway's first "General World Model," introduced in December 2025. GWM-1 builds on Gen-4.5, generates video frame by frame, and accepts camera movements, robot commands or audio as controls. Runway first discussed the real-time approach in March with Runway Characters, and a few weeks before this research it showed Solaris, a system that uses Gen-4.5 to generate user interfaces frame by frame, responding to clicks or voice input. In March, Runway also showed a research preview of a real-time model built with Nvidia and shown at Nvidia's GTC conference; it runs on the Vera Rubin platform and is designed to deliver the first frame in under 100 milliseconds. Runway has not announced a timeline for when a real-time model would be available.

According to Runway, real-time generation closes the gap between an idea and its execution: with instant feedback, users would spend most of their time steering a video rather than waiting for one. Runway also argues it lowers costs, since faster models use less GPU time, and says the cost per output at a given quality level determines which applications make economic sense; instant generation would lower that threshold and make previously unprofitable applications viable. Runway describes small visual errors compounding into major distortions as the central problem with LLM-based approaches to video, since a video model builds each frame on the previous one and cannot correct itself mid-sentence the way a text model can. It addresses this by training the model on its own outputs rather than only on error-free inputs, teaching it to correct its own deviations instead of amplifying them. Runway says real-time generation shifts the compute load from training to use: the model must produce each frame fast enough to keep up with playback while running on hardware shared by several sessions at once.

Startup Decart used a similar approach for its real-time model MirageLSD, deliberately exposing it to flawed or distorted images during training. Google DeepMind says its world model Genie 3 keeps interactive worlds consistent for several minutes at 24 frames per second in 720p. Runway sees interactive applications, such as education, gaming and robotics, as the biggest long-term use case for AI-generated media, since these need video that responds as quickly as the person watching it. Runway previously introduced GWM Robotics, a variant of GWM-1 that generates synthetic training data for robots. Waymo is taking a similar approach with the Waymo World Model, based on Genie 3 and adapted for road traffic; according to Waymo, this lets it simulate situations its fleet has never observed, such as an encounter with an elephant, a tornado or a flooded residential neighborhood, and the Waymo Driver travels billions of miles in virtual worlds before encountering scenarios on public roads.

Key facts

  • Runway is developing real-time video generation that streams and responds to prompts as users describe them, instead of the current wait-then-review workflow.
  • The approach builds on GWM-1, Runway's "General World Model" introduced in December 2025, which builds on Gen-4.5 and generates video frame by frame.
  • A research preview built with Nvidia, shown at GTC in March, runs on the Vera Rubin platform and targets a first frame in under 100 milliseconds; Runway has not announced a release timeline.
  • Runway trains the model on its own outputs, not just error-free inputs, to keep small per-frame errors from compounding into major distortions, the same approach Decart used for MirageLSD.
  • Comparable world models include Google DeepMind's Genie 3, which keeps interactive worlds consistent for several minutes at 24 frames per second in 720p, and the Waymo World Model, which Waymo uses to simulate rare scenarios such as an encounter with an elephant or a flooded street.

Why it matters

Today's video generation tools make users wait seconds or minutes for a finished clip, and any mistake means starting over. Runway argues that streaming generation, responding as a user prompts it, would let people spend their time steering a video rather than waiting for one, and that faster models cost less to run, which could make applications that are not economically viable today worth building.

Who it affects

Video creators and anyone building interactive media are the direct audience, but Runway frames robotics and autonomous vehicles as the biggest long-term use case: it has already built GWM Robotics, a GWM-1 variant that generates synthetic training data for robots. Waymo is a named comparison point, using a Genie 3-based world model to simulate road scenarios its fleet has never seen. Decart, with MirageLSD, and Google DeepMind, with Genie 3, are working on similar real-time or interactive world models.

How to use it

There is nothing to use yet. What exists are previews: Runway Characters (shown in March), Solaris (a Gen-4.5-based system that generates user interfaces frame by frame in response to clicks or voice), and a research preview built with Nvidia for its Vera Rubin platform, targeting a first frame in under 100 milliseconds. Runway has not announced when, or whether, a real-time product will ship.

How solid is it

Every claim in this story, the cost argument, the training method and the use cases, comes from Runway itself; the piece reports the company's research and framing rather than independent testing or benchmarks. The external reference points are secondhand too: Google DeepMind's claim about Genie 3 running at 24 frames per second in 720p, and Waymo's claim about its own simulation mileage, are both companies describing their own systems.

Risks and caveats

Runway gives no concrete cost or GPU-time figures for the savings it claims, only that faster models use less GPU time. It gives no timeline for when a real-time model would be available. The comparison to Genie 3's "several minutes" of consistency is not matched against how long Runway's own model can stay consistent, so the two are not directly comparable on that measure.