CubePart Enables Open-Vocabulary, Part-Controllable 3D Assets for Games
CubePart introduces a two-stage diffusion framework that turns a global prompt and an open-ended part schema into a set of labeled, functional 3D meshes ready for game engines. Stage 1 learns the object’s overall shape from a large text-to-3D corpus, while Stage 2 splits it into per-part meshes (e.g., cab, wheels, headlights) with cross-part attention to ensure proper alignment. Trained on a vast VLM-annotated dataset, the system can decompose existing assets, integrates with animation and scripting workflows, and targets Roblox Studio, with ongoing work to improve deformation and reduce part overlap. The work is open-source and will be presented at SIGGRAPH later this year.