P3D-Bench

P3D-Bench is a benchmark for parametric 3D generation that tests whether multimodal LLMs can generate precise parametric 3D programs and coherent assemblies from text or images. Unlike a mesh, a parametric program exposes explicit dimensions, construction operations, and part relations, so the benchmark can check whether a model recovers a design structure rather than only its appearance.

The benchmark covers three task families, Text-to-3D, Image-to-3D, and Assembly-3D, and scores each output for executability, geometric fidelity, topology, text-grounded constraints, multiview semantic alignment, and part-level structure. The evaluation runs 400 text cases, 400 image cases, and 203 annotated assemblies, and the paper reports assemblies as the hardest setting, where models still fail to compose multiple parts into a coherent structure.

A demo split of 3 cases per task ships in the repo for a zero-setup smoke test, while the full split is published on HuggingFace as UID lists and P3D-derived annotations, materialized from local copies of the Fusion 360 Gallery and Text2CAD sources. The benchmark code is MIT licensed, while the dataset follows its upstream sources and is for non-commercial research use only, with attribution.