Trained to Imagine
Rendering
Creative Intelligence
AI Video / Image Models

Runway introduces GWM Worlds 2, a real-time interactive world model

Source: Runway ·

Runway's research preview streams continuous 720p video at 24fps with 48kHz audio, with no preset session length, steered by text actions and continuous camera movement. Runway lists its own limits: detail degrades on fast camera turns and long-term memory is imperfect.

TTI Takeaway

24fps and 48kHz are film and post numbers, not gaming ones — Runway is pitching this at people who deliver to picture. The stated weak spot, fast camera moves, is the one that matters most for action, so treat it as a tool for locations and mood today, not for coverage.

Analysis

GWM Worlds 2 is the clearest sign yet that Runway is building its world model for people who work to picture, not only for players. The spec sheet reads like a delivery document: continuous 720p at 24 frames per second with 48 kHz audio, generated from a single audio-video foundation model. Those are not arbitrary choices. They are the frame rate and sample rate a cutting room expects, and they say a lot about who Runway thinks will pay for this.

The bigger change from GWM-1, which Runway introduced in December, is control. A session has no preset length and keeps generating as long as inputs arrive. Those inputs come through what Runway calls a WorldPrompt: a genesis prompt that defines the scene, subjects and rules, a first frame, a timestamped stream of text actions that can overlap, and a per-frame camera stream of translation and rotation. Camera and subject are steered independently, multiple roles such as player and director can each have their own actions, and characters can speak with controllable voice, tone and language, with lip movement generated to match. Runway lists filmmaking and advertising under ahead-of-time use, alongside interactive film, games and robotics.

That ahead-of-time mode is the one that matters for production, and Runway says it currently produces better quality than the real-time demo. In practice the WorldPrompt structure looks a lot like a previs brief: a scene description, a starting frame, a beat sheet of actions with timecodes, and a camera move. A director could block a scene, try a different camera path, and play it again without regenerating everything from scratch. For location scouting and mood exploration, that is a real step from prompting isolated clips.

Runway is unusually direct about the limits. Real-time generation trades fidelity for speed. Quick camera rotations degrade detail, texture and geometry, which rules out whip pans and most action coverage. Long-term memory is imperfect, there are no image references beyond the first frame or a prefilled video, and free-form control may need an external harness to track the state of the world. The company’s own line is that “real-time video generation is still in its earliest stages.” The addressable-memory work from NVIDIA and Princeton in August targets exactly the memory problem, and Genie 4’s two-minute persistence in October shows how quickly the bar is moving.

What to watch is whether Runway closes the reference gap. Productions do not start from a blank genesis prompt; they start from a location scan, a production design pack and a cast. Until a world model can be anchored to approved reference throughout a session, not just at frame one, output stays exploratory rather than deliverable. The direction is clear, though, and it follows the $315 million raise in February that Runway explicitly tied to world models. The tools are converging on something a previs team will recognise.