Recently, Roblox announced the launch and open-sourcing of Cube3D, its first foundational AI model for generating 3D objects. According to Roblox's press release, Cube3D aims to improve 3D creation efficiency, allowing developers to quickly explore creative directions and boost productivity.
Roblox Website
The model's innovation lies in its training method. Roblox explains that they drew inspiration from text prediction models, creating a technique that tokenizes 3D objects. Cube3D is trained to predict the next shape token, thereby constructing complete 3D objects.
Regarding training data, Roblox spokesperson Samantha Spielman stated that the model used "a combination of licensed and publicly available datasets, as well as experiential data from the Roblox ecosystem."
Roblox also revealed future plans for Cube3D, stating that the model will eventually become a multi-modal model capable of accepting various input types such as text, images, and videos, and will integrate with Roblox's existing AI creation tools. Particularly noteworthy is that future versions will be able to generate 3D objects using images as input.