Common Sense Machines Unveils AI Platform for Real-Time Level Generation in Games

CSM’s latest 3D AI tools aim to turn simple text prompts, images, and sketches into full game levels in near real time — and they’re built to slot directly into the engines PC developers already use.

On March 15, 2025, Common Sense Machines (CSM.ai) quietly started showing something level designers have been day-dreaming about for years: you describe a space in natural language, drop in a few reference images or sketches, and an AI system builds a playable 3D level around those ideas in minutes instead of days.

The new platform builds on CSM’s existing Cube toolset — which already converts 2D inputs into production-ready 3D meshes and environments — and runs on Google Cloud infrastructure. In practice, it’s pitched as a way to prototype and iterate FPS and RPG maps far faster while keeping human designers firmly in control.

From prompt to playable level

CSM’s core pitch hasn’t changed: “Transform images, text, and sketches into game-ready 3D assets and worlds.” The new level-generation layer essentially strings those capabilities together into a single workflow aimed squarely at game engines.

A typical flow looks like this:

1. Start with intent, not geometry.
   A designer writes a short prompt — for example, “tight, vertical sci-fi research station carved into an icy cliff, three main combat lanes, upstairs overlook for snipers, and a risky mid-map power position.”

2. Add reference material.
   Concept art, mood-board images, or rough grey-box blockouts can be attached so the AI has a sense of scale and style.

3. AI blocks out the level.
   Using its 3D generative models, CSM assembles a navigable layout: rooms, corridors, cover objects, LOS breaks, and traversal routes. It’s more like an automatically built “whitebox” than a finished level — enough to run around in and test.

4. Iterate in real time.
   Designers can feed back edits (“open up this lane,” “add more vertical flanks,” “push this spawn farther back”) as text instructions. The system updates geometry and prop placement while preserving existing sightlines and chokepoints where possible.

5. Export to your engine.
   Once a prototype feels right, the level can be exported and polished in Unity, Unreal, or a studio’s own in-house tools, with meshes and materials already organized into engine-friendly formats.

The idea is not to replace level designers, but to move their time away from repetitive block-out work and toward tuning flow, pacing, and encounter design. You still decide where the sniper nests go — you just don’t have to spend hours dragging walls and doorways into place for every iteration.

Built on CSM’s hybrid 3D engine – and Google Cloud

Under the hood, the level platform is powered by the same tech stack CSM has been building for the last few years:

• Multi-modal 3D generation.
  CSM’s models handle text, sketches, and single images as inputs, then output 3D meshes, Gaussian splats, and animated assets that can be dropped directly into a scene.

• An AI “world renderer.”
  Instead of treating assets and environments separately, CSM runs them through a diffusion-based renderer that fuses traditional graphics with neural rendering. That’s what allows quick restyling of a level — think “same layout, now neon-soaked cyberpunk” — without starting over from scratch.

• Engine-ready workflows.
  On its public site, CSM already advertises dedicated workflows for Unity, Unreal, Blender, and CAD, including “Image to Kit,” “Chat to 3D,” and “Text → Images → 3D → Environment.” The new level tools sit on top of those, packaging them as a more coherent “world-building” experience rather than a pile of separate features.

On the infrastructure side, CSM is leaning on its partnership with Google Cloud. The company’s generative models are available through Google’s Vertex AI Model Garden, and Google has been promoting CSM as part of its broader push toward “living games” — titles that refresh content frequently instead of shipping static levels and calling it a day.

For studios, that matters because it means they can:

• Run CSM’s models inside their own Google Cloud projects, avoiding messy data-sharing questions.
• Scale inference up or down based on production sprints.
• Potentially plug level-generation workflows into the same pipelines they already use for backend services, analytics, and dedicated servers.

Why real-time level generation matters for FPS and RPG teams

Most FPS and RPG studios already iterate on levels constantly — whiteboxes, internal playtests, rebuilds, art passes, and late-stage balance tweaks. The pain is that every idea costs a chunk of time: even a “quick” new layout can chew through days of designer and environment-artist hours.

CSM’s approach doesn’t magically solve the hard parts of level design, but it does attack the slowest ones:

• Faster early prototypes.
  Designers can explore five or ten variations of a map in the time it used to take to build one. That’s a big deal when you’re trying to find a layout that supports your movement, abilities, and weapon sandbox.

• Design-driven, not asset-driven.
  Because geometry is generated from intent and reference, teams don’t have to wait for bespoke art sets before testing ideas. You can design around how a space plays, then lock visuals later.

• Richer LiveOps and seasonal events.
  Live-service games depend on fresh spaces – limited-time event areas, seasonal reskins, and small variations on existing maps. A generative level platform could help support those updates without burning out art and level teams.

For players, the long-term promise is more varied and reactive environments. You might see games where co-op dungeons remix themselves based on party composition, or roguelikes where the “tiles” are chunky pieces of authored geometry assembled by AI rather than simple room templates.

Quality, control, and where the AI stops

If you’ve been burned by over-sold AI tools before, the obvious question is: how good is CSM’s output, really?

A few points worth keeping in mind:

• CSM’s Cube models have scored highly on independent 3D benchmarks, and the company has been publicly showcased by Google Cloud and other partners for its 3D generation work.
• The level platform still outputs editable scenes. Designers can move walls, delete props, or swap materials just as they would in a traditionally built level.
• Studios can constrain generation with style guides and asset packs. Instead of a wild mash-up of shapes, you can tell the system “only use our existing sci-fi corridor kit and lighting presets” so new layouts still look on-brand.

In other words, it’s best to think of this as a super-charged block-out assistant: fast, flexible, and trained on a lot of geometry, but still ultimately a tool in the hands of human level designers.

What it means for PC players and hardware

Most of the heavy lifting for CSM’s level generation happens in the cloud during development, not on your gaming PC. So you don’t need a monster rig just because a studio used CSM to design its maps.

That said, the sort of games that stand to gain most from this tech — big, detailed worlds that change frequently — are also the games that push GPUs and high-refresh monitors the hardest:

• If you want to be ready for denser, more dynamic scenes, check out our Nvidia RTX 5090 review for where top-end GPU performance is headed.
• For the kind of fast-moving competitive games that could benefit from AI-assisted map updates, pairing those worlds with a high-refresh OLED like the Dell AW2725Q QD-OLED — or browsing our monitor hub — will matter more than ever.

On the AI side, you can also imagine future combos where dynamic levels meet smarter NPCs. Nvidia’s ACE framework for AI characters, for example, is already targeting richer, more reactive NPC behavior — something we’ve covered in our story on autonomous characters redefining NPC interaction. Pair that with levels that can be re-shaped on demand, and you start to see where “living games” might actually live up to the name.

Availability, pricing, and what’s next

CSM’s real-time level generation platform is rolling out on top of its existing Cube service:

• Access. Studios can contact CSM directly for demos or pilot projects, and individual creators can already experiment with the underlying asset-generation workflows through the Cube web app.
• Pricing. CSM sells its tools on tiered subscriptions for creators and studios, but it hasn’t publicly broken out separate pricing for the new level-generation stack yet. Expect bespoke quotes for larger teams that want dedicated cloud capacity and tighter engine integrations.
• Engine support. The company is focusing on Unity and Unreal first, with export options that match how teams already manage scenes, prefabs, and materials. Other engines and proprietary pipelines will likely rely on generic mesh and scene export.

For now, this is still early-days tech. But if CSM and Google Cloud are right about where game development is headed, the tools designers use today may soon look a lot more like collaborating with a smart level-design assistant than wrestling with geometry from scratch.

Either way, the takeaway for PC players is simple: over the next few years, you’re likely to see more games where maps evolve more quickly, seasons feel less like simple reskins, and smaller teams can ship worlds that once required AAA budgets.

And when that happens, there’s a good chance at least some of those spaces were first sketched out — in seconds — by Common Sense Machines.