CubePart

CubePart is a generative framework from Roblox Research, with Carnegie Mellon University and Stanford University co-authors, for open-vocabulary, part-controllable 3D mesh generation. The user gives a global text prompt and a parts schema, which is an open-ended list of part names. The model returns one mesh per schema entry, and the meshes assemble into a single coherent object. It is presented at SIGGRAPH 2026.

The method has two stages. Stage 1 fine-tunes a vecset-based diffusion transformer on schema-augmented prompts, so that all requested parts appear in the global shape. Stage 2 reuses the Stage 1 weights and adds zero-initialized Cross-Part Attention Residual Blocks at four layers. These let the parts share structural context while each part is decoded under a part-aware prompt. The authors also built a part-labeled dataset of 462K assets and 2.02M parts, using artist segmentations, vision-language models and 3D-aware Set-of-Mark prompting.

The same object can be decomposed at different granularities by changing the schema, from 2 to 8 parts in the project's examples. The authors say the outputs can be loaded into game engines and driven by animation and behavior scripts without manual post-processing. On PartObjaverse-Tiny they report cleaner part boundaries and better geometric fidelity than PartCrafter, PartPacker, HoloPart and OmniPart baselines. A YouTube reviewer says the code has been released and the model is under 10 GB. The provided pages give no repository, weights link or license, so these are unverified here.