Runway releases GWM Worlds 2 with 720p interactive video at 24 fps

• GWM Worlds 2 generates a continuous 720p, 24 fps video stream with 48 kHz audio, and the product announcement says text actions and continuous camera motion steer it while it runs.
• WorldPrompt separates persistent rules and scene context from timestamped events, a structure the WorldPrompt explanation says can cover actions, speech, sound, and camera movement.
• The same model can run as an authored video, a turn-based experience, or a live simulation, and a GWM Worlds 2 demo shows prompt-driven subject and camera control.
• The project remains a research preview, a label used in , while Runway documents fidelity, memory, and reference-image constraints in its research preview.

A desert demonstration lets a player move a survivor while a director changes the sunset and campfire in the same scene. Runway's world-authoring flow turns a short idea such as “third person perspective dirt bike in a snowy landscape” into a first frame, a scene brief, and editable key-bound actions.

On its research page, Runway divides a world into persistent context and a timestamped event stream.

• Persistent context: a genesis prompt for environment, layout, materials, lighting, ambient sound, subjects, attributes, physical laws, and camera perspective.
• Visual anchor: a first frame that grounds the generation.
• Events: free-form actions addressed to a subject or the scene, with start and end times for movement, gestures, object interactions, speech, and sound.
• Camera: per-frame viewpoint translation and rotation.

Actions can overlap. Speech is represented as an action, carrying the line to be said, rather than as a separate audio track.

Runway's research preview describes three ways to use the same WorldPrompt input.

1. Ahead of time: an author, potentially with an LLM, writes the event stream before generation. Runway names filmmaking and advertising as uses.
2. Turn-based: generation pauses for a user decision, then continues after the next action enters the event stream. Runway positions this for visual novels and interactive film.
3. Real-time: the stream continues while actions immediately affect the next generation steps. Runway positions this mode for games and interactive experiences.

The current live demo maps keyboard and mouse inputs to premade text prompts, so W might mean “The character moves forward” and a left click might mean “The character throws a ball.” Runway says its ahead-of-time method currently produces better quality because those prompts can describe more of the newly visible scene.

Runway first finetunes its foundational audio-video model on WorldPrompt, then post-trains that bidirectional model into the autoregressive GWM Worlds 2. The technical description says the change enables continuous generation rather than a fixed-duration clip.

Each generation step conditions on:

• global context, the genesis prompt and first frame;
• current inputs, the camera signal and actions spanning that frame;
• previous generated frames held in a sliding-window cache.

Older frame tokens are evicted from that cache. Video and audio decoders run causally with their own caches for faster decoding.

Runway also shows agents choosing actions for both a character and the environment, such as steering an adventurer while changing a castle's lighting. In the company's two-flag test, an agent instructed to navigate a humanoid robot to red, blue, or red then blue flags completed each task.

The same post frames these sessions as simulated environments for evaluating and training agents across varied worlds.

The demo assigns actions to roles, including player 1, player 2, and director. Different users can control different subjects, or one participant can control the world itself, according to the multiplayer description.

Runway says its demo uses LiveKit to broadcast video and audio, while each client selects a role in the experience.

Sessions begin from presets containing a first frame, a genesis prompt, and possible subject actions bound to keys. Runway's LLM-assisted authoring flow generates those materials from a short description, then leaves the frame, prompt, and bindings editable before play in the authoring demo.

That setup keeps the free-form WorldPrompt underneath a creator-facing layer of named subjects and reusable controls.

Runway lists several current constraints in its research preview:

• Quick camera rotations can degrade details, textures, and geometry.
• Long-term memory remains imperfect.
• Image references work only as the first frame or as prefilled video and audio.
• Free-form control may need an external real-time harness to track world state and generate actions, particularly for interactions such as NPC dialogue.

The post ends with a form for people to get in touch about GWM Worlds 2, requesting business and company details rather than listing public pricing or an open signup.