Google DeepMind opens up access to Project Genie, an AI tool for creating interactive game worlds from text prompts and images.
Starting Thursday, Google AI Ultra subscribers in the US will be able to try out an experimental research prototype powered by a combination of Google’s latest global model Genie 3, image generation model Nano Banana Pro, and Gemini.
The move, announced five months after Genie 3’s research preview, is part of a broader effort by DeepMind to gather user feedback and training data as it races to develop a more capable model of the world.
A world model is an AI system that generates an internal representation of the environment that can be used to predict future outcomes and plan actions. Many AI leaders, including DeepMind, believe that world models are a critical step toward achieving artificial general intelligence (AGI). But in the short term, labs like DeepMind are envisioning go-to-market plans that start with video games and other entertainment and extend to training embodied agents (aka robots) in simulations.
The release of Project Genie by DeepMind comes as global modeling competition begins to heat up. Fei-Fei Li’s World Labs released its first commercial product called Marble late last year. Runway, an AI video generation startup, also recently launched a world model. Former Meta chief scientist Yann LeCun’s startup AMI Labs will also focus on developing global models.
“I think it’s really exciting to be in a place where we have access to more people and more feedback,” Shlomi Fruchter, DeepMind’s research director, told TechCrunch in a video interview, beaming with obvious excitement over the release of Project Genie.
DeepMind researchers TechCrunch spoke to were candid about the experimental nature of the tool. It’s inconsistent, sometimes producing an impressively playable world, and other times producing baffling results that miss the mark. Here’s how it works:
tech crunch event
boston, massachusetts
|
June 23, 2026
You begin “world sketching” by providing text prompts for both the environment and the protagonist, which later allows you to interact with the world from a first-person or third-person perspective. Nano Banana Pro creates images based on your prompts. In theory, Genie could modify that image before using it as a starting point into an interactive world. The fix mostly worked, but the model would sometimes stumble and give me purple hair when I asked for green.
You can also use real-life photos as a baseline for your model to build your world, but this can also be hit or miss. (More on this later.)
Once you’re happy with the image, Project Genie takes a few seconds to create an explorable world. You can also remix existing worlds into new interpretations by building on prompts, or explore curated worlds using the gallery or randomizer tools for inspiration. You can then download videos of the world you just explored.
DeepMind currently only allows 60 seconds of world generation and navigation due to budget and computing constraints. Because Genie 3 is an autoregressive model, it requires a large amount of dedicated computing, and there are hard limits on how much DeepMind can provide to users.
“The reason we limited it to 60 seconds was because we wanted to reach more users,” Fruchter said. “Basically, when you’re using it, you have your own chip somewhere, and it’s exclusive to your session.”
He added that extending beyond 60 seconds reduces the incremental value of the test.
“The environment is interesting, but at some point the level of interaction limits the dynamism of the environment to some extent. Still, we think that’s a limitation that we want to improve on.”
Weirdness works, realism doesn’t.
When I used the model, the safety guardrails were already up and running. It was not possible to generate anything resembling nudity. Nor could it generate a world where you could even remotely smell Disney or other copyrighted material. (In December, Disney filed an injunction against Google, accusing its AI models of infringing copyright by training Disney characters and IP to generate unauthorized content.) Genie couldn’t even generate a world of mermaids exploring an underwater fantasyland or an ice queen with a winter castle.
Still, the demo was very impressive. The first world I built was an attempt to bring my little childhood fantasies to life. Inside, we got to explore a castle in the clouds made of marshmallows, a river of chocolate sauce and a tree made of candy. (Yes, I was a chubby kid.) I asked the model to do it in the style of clay work. And it delivered a world of whimsy that I would have devoured as a child. The castle’s pastel and white spiers and turrets are plump and look delicious enough to tear off chunks and plunge into a chocolate moat. (video above)
That said, Project Genie still has some issues to work out.
The models excelled at creating worlds based on artistic prompts, including watercolors, anime styles, and classic manga aesthetics. However, they tended to fail when it came to photorealistic or cinematic worlds, often making them look more like a video game than a real person in a real environment.
It also didn’t always respond well when working with real photos. I gave it a photo of my office and asked it to create a world based on it, and it created a world with some of the same furniture in my office (a wooden desk, a plant, and a gray sofa) arranged in a different layout. And it felt sterile, digital, and unrealistic.
You input a photo of a desk with a stuffed animal on it, and Project Genie created an animation of the toy moving through space, sometimes reacting to other objects as it passed.
This interactivity is something DeepMind is working on improving. There have been several times where my character has fallen through walls and other solid objects.
When DeepMind first released Genie 3, researchers highlighted how the model’s autoregressive architecture meant it could remember what it generated. So I wanted to test it by going back to a part of the environment that the model had already generated and see if it was the same. In most cases, this model was successful. In one case, we generated a cat exploring yet another desk, but only once did the model generate a second mug when it returned to the right side of the desk.
The part I found most frustrating was how to use the arrows to look around, the spacebar to jump or rise, and the WASD keys to move. I’m not a gamer, so this didn’t come naturally to me, but the keys often didn’t respond or flew in the wrong direction. Trying to walk from one side of a room to a doorway on the other often resulted in a chaotic zigzag motion, similar to trying to steer a shopping cart with a broken wheel.
Fruchter assured me that his team is aware of these shortcomings and reminded me again that Project Genie is an experimental prototype. In the future, he said, the team hopes to increase realism and improve interaction capabilities, such as giving users more control over their actions and environments.
“We don’t think of[Project Genie]as an end-to-end product that people will come back to every day, but we do think there’s already a glimpse of something that’s interesting and unique and not possible any other way,” he said.
