spatiaos/p3d-bench

P3D-Bench is a research benchmark that tests how well AI models can generate 3D objects from written descriptions or images. It was released in June 2026 by researchers at Nanjing University and Envision, along with a public dataset on HuggingFace and evaluation code in this repository. Most 3D benchmarks only check whether an AI produces something that looks roughly correct. P3D-Bench goes further by testing whether AI can write code that produces a parametric 3D model, meaning a design that encodes exact dimensions, construction steps, and how separate parts relate to each other. This matters because a parametric model can be adjusted and rebuilt consistently, rather than just being a fixed shape that cannot easily be changed. The benchmark covers three types of tasks. In Text-to-3D, a model receives a written description and must output a parametric 3D program. In Image-to-3D, it works from a photo of an object. In Assembly-3D, it must combine multiple parts into a single coherent structure. There are 400 text cases, 400 image cases, and 203 hand-annotated assembly cases. Each output is scored across four areas: geometry accuracy (are the dimensions right), topology (does the shape have the correct structure), semantic alignment checked from multiple viewpoints, and part-level accuracy (are the individual pieces correct in number and shape). The evaluation found three consistent patterns across the AI models tested. Assembly tasks are the hardest: models fail to correctly compose multiple parts together. Models can often capture the overall shape of a target object but miss the precise measurements the input specified. Part-level modeling is the weakest point across all models, with consistent failures on both part geometry and part count. To use the benchmark, you install the Python package, supply API keys for whichever AI model you want to test, download the dataset from HuggingFace, and run evaluations from the command line. Flags let you independently choose the task type, 3D output format, and which scoring metric to apply. Optional dependency groups handle the heavier geometry, rendering, and CAD libraries needed for full scoring.