RareSense/Nova3D

Editable, part-aware 3D generation from text or reference images. Open-source client for nova3d.xyz.

What it solves

Nova3D addresses the limitations of traditional AI 3D generators that produce "baked" or fused meshes. While standard diffusion-based tools create single, opaque objects that are difficult to edit or animate, Nova3D generates 3D assets as executable programs. This ensures that every object consists of named, separately addressable parts with a real assembly hierarchy, making them natively inspectable, editable, and animatable.

How it works

Instead of emitting a mesh directly, the pipeline generates Blender-native Python code. This code is then compiled into a structured GLB file. By using Blender's scene graph as its native representation, the system can leverage Blender's full modifier suite—including subdivision, sculpting, and booleans—to create both precise mechanical parts and organic shapes. The system is model-agnostic, allowing users to swap between different LLM providers (such as OpenAI, Anthropic, or Gemini) to write the construction code.

Who it’s for

It is designed for 3D artists, developers, and designers who need structured, production-ready assets that can be modified, rigged, or retextured at the part level rather than regenerating the entire model from scratch.

Highlights

  • Code-native generation: Assets are created as executable construction procedures rather than static meshes.
  • Assembly hierarchy: Generates labeled assembly trees with joint pivots at articulation points for easy rigging.
  • Part-aware structure: Every component is a separate, named, watertight mesh, allowing for surgical edits and independent retexturing.
  • Automatic UV mapping: Each part receives its own UV chart, auto-packed into a clean, non-overlapping atlas.
  • LLM Agnostic: Works with various LLM providers via a generation harness.