Picture describing a rainy street corner in a sentence, then stepping into it. The rain keeps falling, the street sweeper keeps sweeping, and you can walk around, change the weather, or hand a line of dialogue to a stranger. That’s not a game engine doing the work. It’s Runway WorldPrompt, the input format behind GWM Worlds 2, and it might be the strangest new creative tool of the year.
Text to image already feels normal. Text to an explorable world is a different beast entirely. Here’s what it actually is, minus the research-paper fog.
What a world model actually is
Most AI video tools generate a clip. You prompt, you get ten seconds, the video ends. A world model works differently: it generates a place that keeps going, frame by frame, while you interact with it.
Runway’s GWM Worlds 2 is a research preview of exactly that. It streams continuous 720p video at 24 frames per second, with generated audio at 48,000 Hz, all produced live as you explore. The world has no fixed script and no preset length. You steer with typed actions (“the campfire flares,” “she walks toward the pier”) and the simulation keeps rendering the consequences.
If you’ve ever bounced off game development because engines like Unity demand months of learning, this is the opposite promise. You describe. It persists. That distinction is why people in film pre-visualization, game prototyping, and robotics research are all watching the same demo.
What WorldPrompt actually is
So how do you describe a whole world without writing an essay? Runway WorldPrompt splits the problem in two.
First, the genesis prompt. This is the persistent context: the scene (environment, layout, lighting, ambient sound), the subjects who live in it (with attributes and even voices), and the laws. Laws are my favorite part. Gravity, collision behavior, whether a character can fly, where the camera sits. You set them once, and the world obeys for the whole session.
Second, the event stream. These are timestamped actions, each a plain sentence with a start and end time, aimed at a subject or at the scene itself. Multiple actions overlap freely. A sword swing, a spoken reply, and a room flooding can all happen at once, because speech is just another action, carrying its own line of dialogue.
WorldPrompt is the combination: persistent context plus a stream of events. Runway’s team describes it as a way to control every aspect of the world with text. Worth stressing: it’s a prompting format, not a programming language. You can’t write scripts or manage game state the way you would in Minecraft or Roblox. You suggest, the model complies, more or less.
The street scene example, step by step
Runway publishes a concrete example from its dataset, and walking through it makes the whole thing click.
The scene: a wet city pavement, brick buildings, yellow autumn leaves, reflections on the asphalt after rain. Two subjects: a street sweeper in an orange high-visibility vest holding a broom and dustpan, speaking with a male voice, and a woman in a green trench coat with a female voice, walking away down the pavement. The law section pins Earth-like gravity and locks the camera behind the woman in third-person view.
Then the events roll in with timestamps. At 0 seconds, both characters speak, one saying “Just head straight down,” the other replying a few seconds later. Between lines, the sweeper sweeps leaves and points forward. All of it renders live, characters talking over each other, background city ambience included, from a single authored prompt.
That’s Runway WorldPrompt doing its thing. Filmmakers get a previz shot without a crew. Game designers get a playable slice without an engine. And for the rest of us, it’s a new kind of toy with a weird amount of depth.
Three ways to drive a world
The same model supports three working styles, and picking the right one shapes what you’d build.
Ahead-of-time mode suits filmmaking. You author the full event stream up front, optionally with an LLM helping you draft it, then the model renders the whole thing start to finish. Think of it as directing a generated movie.
Turn-based mode fits visual novels and interactive stories. Generation runs until a decision point, you choose the next action, and the world continues from your choice. Good for branching narrative without scripting trees.
Real-time mode is the video game version. The stream never stops, so your reactions need to land within tens of milliseconds, faster than typing. Recurring actions can be bound to keys, and sessions can even run multiplayer, with different people claiming roles like player or director over a live stream.
What you can build with this right now
Honest expectations first: GWM Worlds 2 is a research preview, not a polished product on your desktop. But the practical uses are already visible.
Prototyping game levels is the obvious one. Instead of grey-boxing a level in an engine to test pacing and mood, describe the level and walk through it in minutes. Filmmakers can previsualize shots, camera moves, and lighting before committing a crew to anything. Educators could reconstruct historical places for students to wander. And on the research side, simulated worlds are becoming training grounds for AI agents and robots, which is exactly why this race isn’t just about entertainment.
There’s also the plain fun angle. People are building tiny dioramas, impossible cities, and dream logic scenes purely to explore them. As a creative toy, it delivers.
How it compares to the rest of the field
Runway isn’t alone here, and the competition tells you how fast this corner of AI is moving. Google DeepMind’s Genie 3 also generates 720p interactive worlds at 24 fps. Odyssey-2 Pro and World Labs’ RTFM (Real-Time Frame Model) are in the same race.
| Runway GWM Worlds 2 | DeepMind Genie 3 | World Labs RTFM | Odyssey-2 Pro | |
|---|---|---|---|---|
| Resolution / rate | 720p @ 24 fps | 720p @ 24 fps | Real-time frames | Real-time video |
| Generated audio | Yes, 48 kHz | Not yet | No | No |
| Control format | WorldPrompt (genesis + events) | Prompt + actions | Image-based scenes | Video-first prompts |
| Multiplayer | Yes, via live sessions | No | No | No |
The audio point is bigger than it looks. Synchronized sound, voices included, generated live with the video, is what makes these worlds feel inhabited instead of like muted dioramas. Runway’s bet on a structured WorldPrompt format rather than a single text blob is the other differentiator. Runway itself is reportedly valued at $5.3 billion, so investors are voting early too.
The honest limitations
Now the caveats, because the demos oversell if you let them.
Fast camera motion degrades quality. Long sessions test the model’s memory, and details drift over time; the world forgets things eventually. You get one first frame for visual grounding, so you can’t feed a stack of reference images. Fidelity trades against speed, which is inherent to generating everything live.
And the control is real but fuzzy. Ask for a specific law of physics or a precise character ability and the model interprets rather than obeys. Anyone expecting engine-grade determinism will bounce off. This is a sketching tool with surprising depth, not a replacement for game engines, and the same trade-off I described in our guide to AI video editing applies here too: speed and cheapness on one side, precision on the other.
Where does Runway WorldPrompt go from here? The research page promises iteration, and if the gap to Genie 3 closes on audio, the last big differentiator in the table above flips sides too.
Where to try WorldPrompt yourself
GWM Worlds 2 lives on Runway’s research page, with published examples you can study, including the full street scene breakdown. Runway’s subscription tiers include world generation features, and the company runs a public research blog documenting progress between releases. If you want to get better at spotting what these models produce, our walkthrough on checking whether a video is AI-generated covers the tells.
If you create anything visual, it’s worth an afternoon. Start with a tiny scene: one room, one character, one law of your choosing. Write the genesis prompt, add three events, and see how long the illusion holds. Then push it, add rain, add a second character, break a law, and watch what happens.
The takeaway: Runway WorldPrompt makes world-building a writing exercise instead of an engineering one. Try one small world this week. You’ll either get a useful prototyping tool or the most fun you’ve had with a text box all year.
Details and quotes from Runway’s official GWM Worlds 2 announcement and Latent Space’s interview with Runway’s CTO.
