Nova3D: LLM-powered pipeline creates 3D objects with articulated, functional parts
Swaps diffusion for LLM-written Blender code to generate clean, movable 3D assemblies.
Nova3D tackles the fundamental flaw in current text-to-3D systems—their inability to reason about part composition. Instead of relying on diffusion or point-cloud generation, the pipeline treats an LLM as a structured code compiler. It writes native Blender Python code blocks that target specific nodes in the scene graph, producing objects with separated, articulated parts like hinges, sockets, and rotating assemblies. For example, a generated washing machine has a churning internal drum, and a Boston Dynamics-style robot dog has functional joints. The final export is a clean multi-part GLB with preserved transform nodes and pivot axes.
The stack is intentionally modular: a Flutter frontend with a Three.js viewport for browser-based rendering and node manipulation, and a backend that hits a hosted API by default but is fully model-agnostic for self-hosting. The creator notes that while local models (e.g., open-source LLMs) can handle basic code generation, they often hallucinate Blender's matrix math—rotating objects into the fourth dimension. For reliable results, they recommend Gemini with your own key. The frontend repo is open source at github.com/RareSense/Nova3D, and the creator invites collaboration to rig it with a strong open-source candidate.
- Uses LLMs to write Blender Python code for scene graph nodes, not diffusion or point clouds.
- Exports multi-part GLB files with working pivot axes, hinges, and sockets for articulation.
- Frontend is Flutter + Three.js; backend is LLM-agnostic; best results with Gemini BYOK.
Why It Matters
Bridges the gap between text prompts and real, functional 3D assets, enabling rapid prototyping of mechanical assemblies.