Roblox open-sources a text-to-3D factory, with strings attached

What it does

Cube is Roblox’s bid to build a “foundation model for 3D intelligence” — starting with text-to-shape generation and expanding toward rigging, animation, and behavior scripting. The current release includes Cube 3D v0.5, an auto-regressive model that generates .obj meshes from text prompts, plus CubePart, which decomposes an input mesh into semantically labeled parts that assemble into a coherent object. Both spit out assets meant to drop straight into game engines.

The interesting bit

The v0.5 model was retrained on ~2.8 million additional synthetic 3D assets with a doubled VQ-VAE latent length (512 → 1024), and it handles bounding-box conditioning — you can specify aspect ratios and watch it stretch a cat into a diagonal abomination to fit. The README is admirably honest about this failure mode, which is rarer than you’d think in corporate AI releases.

Key highlights

• Auto-regressive text-to-3D with optional bounding-box conditioning for aspect-ratio control
• CubePart does open-vocabulary, part-controllable mesh decomposition — one mesh per schema element, animation-ready
• Fast inference path exists but demands 24GB VRAM; 16GB works for the slower route
• Apple Silicon MPS backend supported, though --fast-inference is CUDA-only
• Minimalist Python API (Engine / EngineFast) plus CLI; outputs standard .obj

Caveats

• Bounding-box conditioning can produce disconnected components or diagonal distortions when the box conflicts with the prompt
• --fast-inference flag is incompatible with MacOS and may fail on lower-VRAM GPUs
• Mesh simplification requires an optional meshlab dependency; without it, that step is skipped
• Texture generation and scene layout are listed as “upcoming features” — not here yet

Verdict

Worth a look if you’re building procedural asset pipelines or researching 3D generation. Skip it if you need production textures today or are running on a laptop GPU with dreams of fast inference.