RareSense/Nova3D

What it solves

Nova3D addresses the limitations of traditional AI 3D generators that produce "baked" or fused meshes. Standard diffusion-based generators often create single, opaque objects where parts cannot be independently addressed or edited. Nova3D solves this by generating 3D assets as executable construction procedures (code) rather than static meshes, ensuring that every part of a generated object is inspectable, editable, and animatable.

How it works

Instead of emitting a fused mesh, the pipeline generates Blender-native Python code. This code is then compiled into a structured GLB file. This approach allows the system to create assets with named, separately addressable parts, a real assembly hierarchy, joint pivots for articulation, and PBR materials. The system is model-agnostic, meaning it can use various LLMs (such as those from OpenAI, Anthropic, or Gemini) to write the code, while the pipeline handles the validation and execution.

Who it’s for

Nova3D is designed for creators in games, animation, VFX, and CAD, who need structured, high-fidelity 3D assets that remain actionable after generation—allowing for surgical edits, rigging, and precise dimensional control.

Highlights

• Code-native generation: Assets are treated as programs, making the source of truth the code rather than the mesh.
• Assembly hierarchy: Generates a labeled assembly tree with joint pivots at articulation points, enabling kinematic chains.
• Part-aware construction: Every component is generated as its own named, watertight part from the start, avoiding the need for post-generation segmentation.
• Precision and Organic Flow: Combines the ability to handle precise dimensions (e.g., jewelry CAD) with Blender's modifiers for organic shapes.
• UV-ready: Automatically generates clean, non-overlapping UV atlases grouped by sub-assembly.