Procedura: Agentic 3D Modeling with Procedural Control
Authors: Youtian Lin, Yikang Yang, Zhanpeng Hu, Mengqi Zhou, Feihu Zhang, Xun Cao, Jiaheng Liu, Yao Yao
Submitted: 26 August 2026
Subject: Computer Vision and Pattern Recognition (cs.CV); Graphics (cs.GR)
Project Page: https://spatiaos.github.io/projects/procedura/
Abstract:
Native 3D generators now recover impressive mesh geometry from a single image. However, a dense mesh remains soft where a machined object should be sharp, lacks part decomposition, and exposes no editable parameters. To address these limitations, we explore the paradigm of "3D shape as code," leveraging and scaling the coding abilities of large language models (LLMs) for 3D modeling. We introduce Procedura, a novel agentic 3D modeling framework that represents an object as a procedural assemblyโa parametric program whose named parts are joined by typed, machine-checkable mates. Given a text prompt, the agent plans the object as an assembly graph and writes the program part by part, solving each placement from the mated frames rather than guessing it. A part is admitted only after compile, mate, and connectivity checks pass. A decoupled vision critic then refines the assembly through one diagnosed fix at a time. The same graph also carries per-part materials and a simulator-validated articulation.
We evaluate Procedura on P3D-Bench under its assembly judge and on MechBench-36, our new hard-surface benchmark, using the same judge. On both benchmarks, Procedura outperforms state-of-the-art native 3D generators and all prior 3D-code agents in judged quality, produces the sharpest edges among all evaluated methods, and is the only approach that outputs an editable, part-structured program.
Key Contributions
- Procedural Assembly Representation: We introduce a parametric, part-based representation that supports machine-checkable mates, enabling sharper edges and full editability.
- Agentic Generation Pipeline: An LLM-driven agent plans, writes, and validates the assembly graph step by step, with a vision critic providing iterative refinements.
- Superior Quality and Flexibility: Procedura achieves state-of-the-art quality on both benchmarks while remaining the only method that produces editable, part-structured programs.
Methodology
Procedura operates in three stages:
- Planning: The agent decomposes the target object into an assembly graph, specifying named parts and their typed mates.
- Generation: The agent writes the parametric program part by part, solving placements from mated frames and validating each part via compile, mate, and connectivity checks.
- Refinement: A decoupled vision critic diagnoses issues and applies one fix at a time, iteratively improving the assembly.
Evaluation
We assess Procedura on P3D-Bench (an established assembly-quality benchmark) and MechBench-36 (our hard-surface benchmark). Metrics include judged quality, edge sharpness, and editability. Procedura surpasses native 3D generators and prior 3D-code agents in all key metrics.
Conclusion
Procedura demonstrates that treating 3D modeling as a parametric coding task unlocks a new level of quality and user control. By combining LLM-based generation with procedural validation, we bridge the gap between automatic generation and manual editability, offering a practical solution for professional 3D content creation.
via ArXiv CV