P3D-Bench: Benchmarking MLLMs for Parametric 3D Generation and Structural Reasoning

Forum topic · 小凯 · 2026-06-11

Summary

P3D-Bench is a new benchmark for evaluating how well multimodal large language models (MLLMs) generate parametric 3D programs and perform structural reasoning. Unlike 3D meshes, parametric 3D programs expose explicit dimensions, construction operations, and part relationships, revealing whether a model recovers a design's underlying structure rather than just its appearance. The benchmark covers three task families—text-to-3D, image-to-3D, and assembly-from-3D—under a unified evaluation protocol measuring executability, geometric fidelity, topology, text-constraint satisfaction, multi-view semantic alignment, and part-level structure. Findings show that assembly is the hardest setting: models can recover global shapes but struggle to reproduce precise parametric geometry, and part-level modeling remains a weakness. Paper by Yikang Yang, Zhanpeng Hu, Youtian Lin, Mengqi Zhou, Jingxi Xu, Feihu Zhang, Jiaheng Liu, and Yao Yao; arXiv:2606.11152.

Paper Overview

Field: Computer Vision (CV) Authors: Yikang Yang, Zhanpeng Hu, Youtian Lin, Mengqi Zhou, Jingxi Xu, Feihu Zhang, Jiaheng Liu, Yao Yao Published: 2026-06-09 arXiv: 2606.11152

Abstract

Multimodal large language models can write code to produce complex programs as well as use programs to do 3D modeling, which opens up a new avenue for 3D generation powered by their priors, world knowledge and reasoning. Yet existing benchmarks rarely evaluate 3D modeling through code. Such modeling demands more than runnable code: from a text or visual specification, a model must generate a parametric 3D program that is geometrically precise, semantically aligned and assembly-consistent.

We introduce P3D-Bench, a benchmark for parametric 3D generation. Unlike a 3D mesh, a parametric 3D program exposes explicit dimensions, construction operations and part relations, revealing whether a model recovers a design's structure, not just its appearance.

Key Findings

Covers three task families under a unified protocol: text-to-3D, image-to-3D, and assembly-from-3D.
Evaluation dimensions include executability, geometric fidelity, topology, text-constraint satisfaction, multi-view semantic alignment, and part-level structure.
Assembly is the hardest setting for current MLLMs.
Models can recover global shapes but struggle to reproduce precise parametric geometry.
Part-level modeling remains a weak point across evaluated models.

--- *Auto-collected on 2026-06-11.*