Procedura: LLM-Written Procedural Assemblies Outperform Native 3D Generators

Procedura reframes 3D modeling as program synthesis: an LLM writes a parametric procedural assembly whose parts are validated by compile, mate, and connectivity checks. Reported on P3D-Bench and the new MechBench-36 benchmark, it produces sharper edges and the only editable, part-structured output among evaluated methods. Evidence is limited to an abstract without released code, judge definitions, or numerical scores.

August 29, 20261 source40% confidence

68Strategic

Strategic score

Evidence strength30

Technical novelty85

Commercial potential70

Procedura marks a departure from direct mesh generation: instead of regressing dense geometry from an image or text, it prompts an LLM to write a procedural assembly—a parametric program with named parts and typed mates—and admits a part only after compile, mate, and connectivity checks pass. The core novelty is the reframing of 3D modeling as constrained code synthesis, where each part's placement is solved from mated frames rather than inferred as free geometry.¹

Context

This reframing moves the critical constraint. Traditional native 3D generators optimize vertex positions, leaving soft edges, no part decomposition, and no editable parameters. Procedura's constraint becomes machine-checkable program semantics: a part is correct if it assembles, mates, and connects. The reported results, that it produces the sharpest edges of any method evaluated and is the only editable, part-structured output, follow naturally from that design.¹

Assessment

As evidence, the abstract claims state-of-the-art judged quality on P3D-Bench under its assembly judge and on a new MechBench-36 benchmark. No numerical scores, no model details, and no ablations are reported. The evaluation rests on an unpublished judge and an unpublished benchmark, which limits independent verification.¹

Caveats

The claims would become convincing if the authors release code, the MechBench-36 dataset, and judge definitions, and if independent evaluations confirm the assembly quality advantage against native generators on standard metrics. Until then, the work is best read as a promising direction with strong internal logic but weak public evidence.¹

Source text

Hugging Face Daily Papers — Procedura: Agentic 3D Modeling with Procedural Control · August 26, 2026

Computer Science > Computer Vision and Pattern Recognition

Title:Procedura: Agentic 3D Modeling with Procedural Control

Abstract:Native 3D generators now recover impressive mesh geometry from a single image. However, a dense mesh stays soft where a machined object should be sharp, it carries no part decomposition, and it exposes no parameter a user could edit. To address this, we explore the paradigm of 3D shape as code, leveraging and scaling the coding ability of an LLM for 3D modeling. We introduce Procedura, a novel 3D modeling agent framework that writes an object as a procedural assembly, a parametric program whose named parts are joined by typed, machine-checkable mates. From 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, and admitting a part only once compile, mate, and connectivity checks pass. A decoupled vision critic then refines the assembly one diagnosed fix at a time. Moreover, the same graph carries per-part materials and a simulator-validated articulation. We evaluate on P3D-Bench under its assembly judge, and with the same judge on MechBench-36, our hard-surface benchmark. On both, Procedura outperforms state-of-the-art native 3D generators and every prior 3D-code agent on judged quality, produces the sharpest edges of any method we evaluate, and is the only one whose output is an editable, part-structured program.

Bookmark

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

The phrase 'typed, machine-checkable mates' is not decorative: it means the interface between two parts (e.g., a pin and a hole) is a formal type with clear semantic rules. This is the core enabling mechanism that distinguishes Procedura from text-to-3D agents that only output raw meshes or generic code.

The 'assembly judge' is an evaluator that supposedly measures whether parts fit together, yet no definition is provided. A judge designed by the authors to reward their own output format can make 'judged quality' incomparable to mesh-based metrics like Chamfer distance.

The 'decoupled vision critic refines the assembly one diagnosed fix at a time' implies a local repair loop. The unstated assumption is that a single global structure can be reached from a flawed assembly by local edits, which is not guaranteed for interlocking mate graphs.

The source text above remains the property of its publisher and is reproduced for commentary.

↑ Back to top