Nova3D Turns 3D AI Assets Into Programmable Blender Code

Researchers have introduced Nova3D, a system that generates 3D assets as executable Blender Python code rather than opaque, fused meshes. This shift transforms AI-generated 3D content from static visual shells into programmable components that possess named parts and assembly hierarchies from the moment of creation.

Unlike traditional generative AI tools that produce a single mesh, Nova3D treats the resulting binary glTF (GLB) file as a compiled artifact of an underlying program. This representational change turns a 3D object from an opaque surface into a programmable asset that downstream systems can inspect, measure, edit, and animate with surgical precision.

High-Precision Performance

In evaluations using the Nova3D-Bench, which consisted of 54 items, the system produced a valid executable program and artifact for 100% of the test cases. The system demonstrated a significant leap in following technical specifications, satisfying 51 out of 52 prompt-stated numeric and count constraints. In comparison, the best-performing baseline achieved only 11 out of 52.

Beyond simple geometry, Nova3D showed a superior ability to handle articulation. The system successfully articulated 59 joints across 12 assets with 98.3% geometric validity. Notably, all tested baselines failed to expose any native joints, highlighting a critical gap in current 3D generative capabilities.

Bridging the Gap Between Art and CAD

Most current 3D generative AI relies on diffusion to create visually impressive but "fused" meshes. While these look realistic, they are notoriously difficult to edit or animate because they lack internal semantic structure. Conversely, CAD systems provide rigorous structure but often struggle to produce organic shapes.

Nova3D bridges this divide by using the Blender API as its native representation. By combining the precision of code with the flexibility of a full 3D suite, the system allows for the creation of complex, interactive worlds. This approach has already seen practical application in jewelry CAD via FormaNova, where assets must be measured accurately for real-world manufacturing.

A Model-Agnostic Future

To ensure flexibility, the researchers built Nova3D to be model-agnostic. Through a dedicated generation harness, users can swap between different large language model (LLM) providers, including OpenAI, Anthropic, and Gemini, depending on their specific needs.

As the industry moves toward more interactive 3D environments, the ability to generate assets that are pre-rigged for animation and programmatically modifiable removes the need for extensive manual post-processing. The focus now shifts to how these programmable assets will integrate into larger game engines and industrial design pipelines.