Code-native generation of highly programmable 3D assets (2026)

[Submitted connected 22 Jul 2026]

View PDF HTML (experimental)

Abstract:Current 3D generative models mostly nutrient a last surface: a visually beardown but mostly opaque mesh. Interactive 3D worlds request much than a surface. They request named parts, an assembly hierarchy, measurable constraints, section edit handles, and joints for articulation. We coming Nova3D, a strategy that generates 3D assets arsenic executable Blender root code; the compiled mesh, a binary glTF (GLB), is treated arsenic the artifact, not the asset. Because the output is simply a program, semantic handles beryllium astatine procreation clip alternatively than being recovered afterward by segmentation aliases rigging. We measure connected Nova3D-Bench, a frozen, spec-grounded benchmark of 54 items crossed six domains and 3 trouble levels pinch matter and image inputs, against eleven baselines successful 4 families (mesh-native, part-structured, code-native, and CAD) positive a same-LLM ablation. Nova3D produces an executable programme and a valid artifact for 54/54 items. Every plus exposes named parts organized successful a parent-child assembly tree; nary mesh-native, CAD, aliases segmentation baseline exposes either. It satisfies 51/52 prompt-stated numeric and count constraints (best baseline: 11/52), passes 14/18 blinded section edits pinch locality preserved successful 18/18, and articulates 59 joints crossed 12 assets astatine 98.3% geometric validity, wherever each baseline exposes zero autochthonal joints. Its geometry is competitive: it wins the system domains successful a pairwise shape-quality tourney and is 2nd only to the strongest mesh-native model, while conceding texture realism to baked-PBR systems. The cardinal consequence is representational: code-native procreation turns a generated 3D entity from an opaque aboveground into a programmable plus that downstream systems tin inspect, measure, edit, and animate.