Roblox's CubePart generates modular 3D assets ready for game engines
Roblox just unveiled CubePart, a 3D generative model that outputs functional, pre-separated meshes tailored to a developer's defined schema. Instead of spitting out a single solid mesh, it generates accurately labeled parts like spinning wheels and opening doors that drop directly into physics and animation pipelines. It turns generative 3D from pretty digital statues into actual interactive assets.
Anyone who builds games knows that current 3D generators have a major practical catch.
Type in a prompt and you will get a gorgeous sports car in seconds. But bring it into Unreal, Unity, or Roblox Studio, and you realize it is a useless, monolithic paperweight. The wheels are welded to the body, the doors cannot open, and turning on the headlights is impossible without an artist spending half a day manually slicing up the mesh and naming parts.
Roblox is tackling that bottleneck head-on. The company just introduced CubePart, a two-stage 3D diffusion framework designed specifically to output interactive, game-ready assets split into cleanly separated, labeled parts.
Generative meshes that talk to game code
The magic here lies in what Roblox calls a schema, which acts essentially as an API contract between the model and the game engine.
Instead of just feeding the generator a prompt like "a jelly fish themed race car," creators provide a second input: an open-vocabulary parts list. You can ask for a chassis, a gun, exhaust pipes, and four distinct wheels. CubePart then generates each piece as an independent mesh adhering to those exact names, meaning physics constraints, animation rigs, and Lua scripts can hook right into them with zero manual cleanup.
Under the hood, Stage 1 generates the overall shape using an MMDiT backbone paired with a Qwen-VL text encoder, trained on roughly 4.7 million mesh-text pairs and fine-tuned to be schema-aware. Stage 2 refines the geometry into distinct parts.
Teaching a neural net spatial orientation—such as telling a front-left wheel apart from a rear-right wheel—required massive data. Roblox built an automated vision-language pipeline to label existing 3D assets, amassing a dataset of over 460,000 models and 2.02 million parts, more than 11 times larger than previous public benchmarks.
The framework can also decompose legacy artist meshes into new schemas for retrofitting existing assets. Roblox published the paper on arXiv, updated its open-source Cube repository with part control, and plans to roll the tool straight into Roblox Studio.
Why it matters
CubePart marks the moment generative 3D starts shifting from digital figurines into functional game mechanics.
For nearly two years, the AI 3D space focused heavily on surface fidelity, neat quads, and rapid texturing. Yet for interactive media and robotics simulation, an unarticulated mesh is essentially dead on arrival. Roblox bypassed the vanity metrics and solved an unglamorous engineering problem: how to make AI outputs actually obey gameplay scripts.
For independent developers and UGC creators, this is a massive productivity leap. Prototyping an interactive vehicle or customizable prop previously required manual asset breakdown that killed any speed advantages gained from AI generation. Lowering that barrier could speed up game prototyping by orders of magnitude.
That said, the technology is still young. Roblox explicitly notes that CubePart currently only handles rigid-body decomposition. Organic characters with skinned vertex weights are still in the works, parts can still overlap at junctions, and spatial reasoning occasionally gets confused.
Even with those caveats, Roblox's approach is deeply pragmatic. Instead of treating 3D as a static visual, it treats 3D as interactive code.
Sources (1)
- about.roblox.com(opens in a new tab)Originalabout.roblox.com