The Demo That Changed the Verb
Google DeepMind’s Genie 3 is the first real-time interactive world model that actually earns the name. Type a description and it generates a photorealistic, explorable world, running at 720p and 24 frames per second, staying coherent for minutes of continuous interaction. You do not watch the output. You move through it, and the world holds together as you go.
That is a different product than video generation, even when the frames look similar. A video model answers the question of what a place looks like. A world model answers the question of what it is like to be there. For anyone who builds brand worlds for a living, that distinction is not academic. It is the whole job changing shape.
A Field, Not a Fluke
Genie 3 is the headline, but it is not alone. A wave of competing systems, Oasis, GameNGen, Mineworld, Matrix-Game, is demonstrating real-time open environments built on streaming diffusion, with physics that emerge from the model rather than from a hand-coded engine. Matrix-Game 3.0 adds long-horizon memory, so the world remembers what you did in it minutes ago.
When one lab shows a trick, it is a demo. When five labs converge on the same capability within a year, it is a direction. Real-time world generation is now a direction, and the pace looks a lot like AI video did two years before it landed in every campaign budget.
“The render is no longer a film. It is a place. And a place asks something different of the people who make it.”
From Watching to Walking
The shift underneath all of this is from passive generation to active interaction. Every AI video tool to date produces something an audience receives. World models produce something an audience inhabits. The frame stops being a window and becomes a door, and the audience stops being a viewer and becomes a visitor.
That changes the creative brief in a fundamental way. A film has a beginning, an edit, and a running time you control. A place has sightlines, paths, and moments people find on their own. You stop authoring a sequence and start authoring a space, and you have to trust the visitor to walk it in an order you did not choose.
We Have Been Building Worlds Already
Our studio already builds photoreal brand worlds in Unreal Engine 5 for campaigns. We took SeaWorld into deep-sea cinematics and built cosmic worlds for New Era, and in both cases the world was the asset: art-directed, lit, and staged like a set, then shot for the film. World models are the next step on the same road. Today you art-direct a world and shoot it. Soon the audience will walk through it themselves.
The encouraging part is how much of the craft carries over. Art direction, lighting, staging, restraint: a generated world without those is just square footage. The teams who already know how to make a place feel like a brand will be the ones who make these models sing, the same way cinematographers made early AI video watchable.
What Brands Can Do With This Now
The near-term uses are more practical than the demos suggest. Explorable product environments, where a customer wanders the world a product implies instead of watching a thirty-second tour of it. Virtual location scouting in minutes instead of weeks, generating and walking a dozen candidate settings before lunch. Live backdrops for virtual production, generated to the brief rather than licensed from a library.
And the one we find most interesting: campaign microsites where the hero film becomes a hero place. The sixty-second spot introduces the world, and the site lets people step into it. The film and the place share one art direction, one grade, one feeling. That is a campaign architecture nobody could sell a year ago.
“Brands spent a century buying attention inside other people’s worlds. The next decade is about building worlds worth visiting.”
The Honest Limits
Now the caveats, because the demos hide them. Coherence holds for minutes, not hours. Resolution sits at 720p, which is fine for a laptop window and wrong for a stadium screen. And there is no persistent state across sessions yet: leave the world and come back, and it will not remember you were there.
So we treat 2026 world models the way we treated early AI video: as pre-viz and prototype tools while the fidelity climbs. They are already good enough to scout, to pitch, and to prototype a place before you commit an Unreal build to it. They are not yet good enough to be the flagship experience. That line will move, and probably faster than anyone budgets for.
Worlds Worth Visiting
Step back far enough and the pattern is old. Advertising has always been a tenant, renting thirty seconds inside someone else’s show, someone else’s feed, someone else’s game. World models offer brands the landlord’s side of the deal for the first time at a cost that is not a theme park.
The bar, though, is brutal. Nobody has to visit your world, so it has to be worth the trip. That is a craft problem, not a compute problem, and it is exactly the kind of problem our studio likes. We spent the last few years learning to build places people want to look at. The next few are about building places people want to stay.