Guides
Roblox Cube 3D: what it is, how to use it, and when to use another generator
Roblox Cube 3D explained: how it generates meshes, using /generate_mesh in Studio, the GenerationService API, the open-source model and its limits.
3 min read

Quick answer
What is Roblox Cube 3D?
It's Roblox's own generative model for 3D objects, released in March 2025. Cube 3D turns text, and now images, into meshes by predicting 3D "shape tokens" the way a language model predicts words. It powers mesh generation in Studio's Assistant and the in-game GenerationService API, and an open-source version is on GitHub.
What Cube is
Roblox introduced Cube on 17 March 2025 as its "core generative AI system for 3D and 4D". The first piece, Cube 3D, generates a 3D object from a text prompt.
Most 3D generators reconstruct shapes from generated images. Roblox took a different route. It tokenises 3D shapes and trains a transformer to predict the next shape token, the way a language model predicts the next word. Roblox says this lets it train on native 3D data rather than images. The "4D" in its ambitions means objects that function: cars that drive, doors that open.
A short timeline
- March 2025. Cube 3D launches.
/generatearrives in Studio's Assistant as a beta, and the model is open-sourced on GitHub and Hugging Face. - July 2025. v0.5 adds bounding-box conditioning and was trained on millions more assets. The in-experience rate limit rises from 5 to 30 generations a minute.
- March 2026. Assistant mesh generation gains batch creation, bounding boxes and a maximum-triangle setting.
- May 2026. CubePart, which generates an object as separate named parts that can be scripted, joins the open-source repository.
Using it in Studio
In Assistant, type /generate_mesh followed by a prompt, or just ask, for example "make a futuristic crate". From Roblox's Assistant guide:
- the input is text or an image, but not both in one request;
- select a Part first to use it as a bounding box, and the mesh fits inside it;
- set a maximum triangle count (default 10,000). Lower values give more faceted, low-poly results.
The mesh lands in your place, textured. Generated content is moderated, and some prompts are refused.
Using it in a game
GenerationService lets server scripts generate meshes for players while the game runs, for example "a red car of the future" in a vehicle-builder. Its newer GenerateModelAsync method takes a text prompt, an image or both, plus a schema such as a five-part car chassis, and returns a Model. Read Roblox's generative AI rules first. Player-facing generation has to be disclosed in the Content Maturity questionnaire.
The open-source model
You can run Cube 3D yourself from the Roblox/cube repository. The command-line tool writes an .obj, takes an optional bounding box, and needs a capable GPU, with Apple Silicon supported through PyTorch's MPS backend. This is for researchers and tinkerers, because the Studio integration is far easier for building games.
When to use something else
Cube in Studio is the fastest way to get a prop into a place without leaving Studio. An external generator is the better fit when:
- you want to draw it. A sketch fixes the exact silhouette. See turn a drawing into a Roblox model;
- you want to review before import, comparing versions, checking topology and keeping the files;
- the same assets go elsewhere, to Blender, Unity or the web.
Vi3W does those three. You give it a prompt, photo or sketch, it retopologises to low poly, and it exports GLB, FBX or OBJ. Whichever generator you use, the import rules are the same (how to make Roblox models with AI). Try Vi3W free.
Last reviewed

