Cube 3D is a generative AI model from Roblox that creates 3D objects from text. It is the first part of Cube, Roblox's foundation model for 3D intelligence; a foundation model is a large model trained on broad data that can serve many tasks. Roblox's longer goal is a model that helps developers produce whole Roblox experiences, from objects and scenes to rigged characters and scripts.
Cube 3D has two parts. A shape tokenizer, a VQ-VAE (a neural network that compresses data into entries of a learned codebook), turns a 3D shape into a sequence of discrete tokens and back. An autoregressive text-to-shape model then predicts these tokens one after another from a text prompt, the way a language model predicts words. The output is a mesh saved as an .obj file. Version 0.5 adds bounding box conditioning, which sets the overall proportions of the object.
The model weights are on Hugging Face and the inference code is on GitHub, free to download. The repository's license is a research-only RAIL license; the Hugging Face page for version 0.5 lists an OpenRAIL license.