3DCode: Procedural 3D Modeling Code Generation Benchmark

3DCode is an open-source reproduction of the 3DCodeBench academic benchmark. It contains datasets for evaluating models that generate Blender 5.0 Python code to procedurally build specific 3D objects. The benchmark includes 212 categories, each with a canonical seed, a reference factory, and two text prompts.

Use Cases

• Benchmarking AI models for procedural 3D modeling based on the described 212 categories.
• Training models to generate Blender Python code based on text prompts.
• Evaluating the execution success of generated 3D modeling code.
• Studying the relationship between natural language instructions and procedural 3D object creation.

Strengths

• Benchmark includes 212 distinct 3D object categories.
• Provides a canonical seed and two text prompts per category for evaluation.
• Specifically targets Blender 5.0 Python code generation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

YipengGao on Hugging Face.

Collection Method

Likely compiled for academic benchmarking of procedural 3D modeling via code.

Freshness

Last updated 2026-05-24 23:47:05; freshness should be verified.