CommunityNSFW

Nova3D generates articulated 3D parts via LLM-compiled Blender code

A new open-source pipeline uses language models to write native Blender Python instead of diffusion meshes, producing multi-part 3D objects with working hinges and swappable components.

May 17, 2026

Nova3D generates articulated 3D parts via LLM-compiled Blender code

Nova3D is an open-source 3D generation tool that treats language models as structured code compilers rather than image generators. Instead of producing monolithic mesh blobs via diffusion, the pipeline writes native Blender Python that targets specific nodes in the scene graph. The result is a multi-part GLB export with preserved transform nodes, working pivot axes, and articulated assemblies — a washing machine with a door that actually swings, a robot dog with jointed legs, a microwave with a separate handle and cavity.

More in Community

Watch: Nova3D generates articulated 3D parts via LLM-compiled Blender code

The distinction matters because current text-to-3D systems regenerate the entire object when you change a single word in the prompt. Nova3D's approach lets you swap a gun's scope or a robot's arm without recompiling the whole model, because the pipeline understands parts as discrete scene-graph entities. The frontend is Flutter with a Three.js viewport for in-browser manipulation; the backend is model-agnostic and defaults to a hosted API, but self-hosting works with any LLM endpoint.

Local models still struggle with the approach. The creator reports hallucinations in Blender's internal matrix math on complex transforms — "correct-looking code that rotates things into the fourth dimension" — and recommends Gemini as a bring-your-own-key fallback for best results. The frontend repo is live on GitHub, and the creator is actively soliciting technical feedback on prompt-to-code pipelines in 3D contexts and looking for collaborators to integrate a reliable open-weight model candidate. Tighter guardrails on transform hallucinations and a curated list of local models that handle Blender's API without inventing quaternion functions would clear the path to a fully local workflow.