Smarter 3D Prop Creation and the Future of Large World Models (LWMs)
Keywords Studios
Partnering on the world’s favourite video games and TV shows. Global reach. Creative impact. End-to-end support.
Welcome to the April edition of Supercharged, where we cover the latest innovative ideas shaping the future of games and entertainment.
This month, our R&D team shares how AI can help speed up 3D prop creation and gives you full access to their GDC talk about Project KARA. No GDC Vault access needed.
Plus, we’ll look at four pioneering approaches to AI world models, each still finding its footing but showing remarkable promise for the future of game development.
Happy reading!
Project KARA is our research & development initiative to remaster the game Detonation Racing by Electric Square with AI infused pipelines. This month’s pipeline spotlight:
🚧 Generative AI-infused 3D prop creation
Props are the unsung heroes of immersive game worlds, guiding players through open spaces and hinting at the culture, history, and mood of each location. They don't just fill empty spots—they help define the visual style and storytelling of in-game environments.
When players praise a game world for feeling “alive” or “believable”, chances are they’re noticing thoughtfully placed lights, rocks, books, banners, broken walls, and more. In racing games, tires, cones, billboards, and pit areas create a vibrant backdrop and enhance the immersive storytelling.
But that deep immersion comes with a production challenge: the more believable the world, the more props you need. Even purely decorative, non-functional props must match the game’s quality and consistency. That’s where smart tech comes in to take on some of the repetitive grind.
Curious how AI can potentially speed up your 3D asset creation without sacrificing artistry?
We’ve remastered a racing game with GAI-infused pipelines. Here is what we learned.
Watch our full GDC 2025 presentation to see real-world results from our R&D initiative Project KARA.
If you’ve been following this project through our previous Supercharged newsletter editions, this is your chance to go deeper. We have the full talk from James Roadley-Battin for you featuring in-game footage, future opportunities, key benefits, and the challenges we’ve encountered along the way. No GDC Vault access is needed.
Watch now! 👇
AI World Models: The Emerging Frontier of Game Development
By Stephen Peacock , Head of Games AI
Creating rich, immersive game worlds has followed the same laborious path for decades. AAA studios invest years—and millions—placing every tree, rock, and building by hand. Each new village or ruin represents months of painstaking work. Artists sketch concepts. Modelers build assets. Lighting specialists create an atmosphere. Level designers integrate it all.
The high cost of world-building limits our ability to deliver the worlds we imagine. It constrains our creative ambitions and restricts the scope of what's possible.
But something interesting is happening. Four pioneering approaches to AI world models have emerged, each still finding its footing but showing remarkable promise.
Microsoft, Google DeepMind, and World Labs have introduced early versions of large world models (LWMs)—AI systems that can generate interactive, explorable spaces with minimal human input. Let's look at these four distinct approaches:
💡 Microsoft's Muse (WHAM) is a generative AI model trained to simulate gameplay by jointly modelling both the game world and human player actions. Trained on over 1 billion frames from Bleeding Edge (roughly 7+ years of gameplay), it can generate accurate-looking gameplay sequences up to a couple of minutes long. Its strength lies in gameplay ideation—designers can visualize new mechanics before implementing them in code.
💡 Microsoft's WHAMM (World and Human Action MaskGIT Model) pushes this approach into real-time interactivity. Using a parallel generation strategy rather than slow autoregressive generation, WHAMM powers a playable Quake II demo running entirely inside the AI model. While it has only 0.9 seconds of memory and some visual limitations, it represents a significant leap—effectively emulating a game without a traditional engine.
💡 Google DeepMind's Genie 2 creates entirely new 3D worlds from text prompts. It offers potentially infinite world variety with emergent physics. However, it has a short memory (about 1 minute) and lacks structured gameplay. Its promise? Dynamic procedural generation at an unprecedented scale.
💡 World Labs' technology transforms 2D images into explorable 3D scenes. It excels at quickly creating spatially persistent environments, though these spaces remain relatively static with minimal interactivity. Its future lies in accelerated level design and virtual world expansion.
What excites me most is the promise: Imagine providing a load of concept art, a few "beauty corners," a rough blocked-out world, and some reference photos—and watching as an almost infinite world unfolds, consistent with your vision but extending far beyond what any team could manually create.
Will these approaches eventually transform world construction completely? Yes, but don’t panic.
When it comes to game worlds, demand will always outstrip supply. Whenever technology allows us to dream bigger and produce more, players raise their expectations. Their appetite for rich, varied, creative, exciting worlds is infinite.
Today's LWMs have fundamental limitations—memory constraints, fidelity issues, and limited controllability; they are just research projects right now. But they're improving at a breath taking pace, and this is just the beginning.
I'm optimistic about what's coming. As these approaches mature, they will help us build worlds more alive and immersive than we can imagine today.
Which of these four approaches to AI LWMs appeals to you most, and why?
Are you seeing gameplay simulation, real-time exploration, prompt-based world creation, or 2D-to-3D transformation as having the most significant impact on your work? Let us know in the comments below or reach out to Stephen directly:
How external development became a critical layer in game development
The video games industry is shifting, and long-term trust is driving the change. Our CEO Bertrand Bodson shared what he's hearing from the top publishers and how studios are rethinking the way games get made.
Superfan apps: the antidote to social media cynicism?
MediaCat recently explored where fan-first brands in music, sports, and gaming should be heading next. Waste Creative ’s Head of Strategy, Christian Perrins, shared his thoughts on what real engagement looks like in the gaming space, and why it pays to know your audience.
Debunking the top 5 Trust & Safety myths in gaming
The Gaming Safety Coalition hosted a myth-busting LinkedIn Live earlier this month. Our Global Head of Trust & Safety, Sharon Fisher , alongside valued coalition members, shared real-world case studies and actionable insights to challenge misconceptions such as "moderation is censorship" and "content moderation is only a concern for large social media platforms."
We hope you found something you love in this month’s edition. If not, we’re here to listen: what other topics would you like to see covered in the coming months? Let us know in the comments below. Or contact us here to discuss your questions, suggestions, and future projects.
If you’d love to receive our upcoming editions in your email inbox, you’re welcome to sign up here.
Please note that the inclusion of references to GAI tools in this newsletter is not an endorsement of the suitability of such GAI tools generally and all such GAI tools should be tested and approved on an individual basis prior to being used.