Roblox Cube
Roblox's research foundation model that generates 3D shapes from text prompts for game assets.
About
Roblox built Cube as a generative foundation model for 3D intelligence, and the open release centers on text-to-shape generation for game assets: a shape tokenizer turns meshes into discrete tokens, and the model generates new shapes from text prompts, with bounding box conditioning added in the July 2025 v0.5 release and part-controllable generation arriving with CubePart in May 2026. Texture generation and scene layout generation sit on the published roadmap. Running inference takes a minimum of 16 GB of VRAM with 24 GB recommended, tested on NVIDIA L40S, H100, and A100 cards as well as Apple Silicon M2 through M4 machines, and installation is a git clone followed by pip install -e with an optional meshlab extra. The catch is licensing: code and weights ship under the CUBE3D Research-Only RAIL-MS license, which limits use to academic or research purposes and forbids commercial products, hosted services, and commercial derivatives. Even so, it remains a rare case of a major game platform publishing its 3D foundation model in the open, and the repository holds about 1.2k GitHub stars.
Should you use Roblox Cube?
Pick it when
Pick Roblox Cube for academic work on tokenized text-to-shape generation, including bounding box conditioning, especially if you want to run on Apple Silicon M2 to M4 or a 16 to 24 GB NVIDIA card.
Look elsewhere when
Skip it for any commercial or hosted use: the Cube3D RAIL-MS license is research only. Texture generation is still on its roadmap, so TRELLIS or Hunyuan3D-2 fit text-to-3D work that needs textures.
Alternatives to Roblox Cube
- TRELLIS
MIT licensed text or image to textured mesh, Gaussians, or radiance fields, commercially usable, at a 12 GB floor.
- Hunyuan3D-2
Text or image to textured assets with commercial rights below 1 million monthly users, outside the EU, UK, and South Korea.
- Shap-E
MIT licensed text-to-3D that runs on 8 GB and gives textured meshes, though at lower fidelity than optimization methods.
- TripoSG
MIT shape generation from images, including sketches, if you can start from a picture instead of text.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Category
- 3D Model Generation
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Intermediate (3/5)
- License
- Cube3D Research-Only RAIL-MS
- Minimum VRAM
- 16 GB
- Added
- Jul 29, 2026