Procedura Turns 3D Generation into Editable Assembly Code

Introduction

Single-image 3D generators have become good at recovering plausible overall geometry. That is not the same as producing a useful mechanical asset. A dense mesh may look convincing while retaining soft edges, no explicit part breakdown, and no meaningful parameters for a designer to change.

Procedura takes a different route: it treats a 3D shape as code. Instead of asking a model to guess a finished mesh in one pass, the framework uses an agent to write a parametric assembly whose components and relationships remain explicit.

What the system does

• Plans an assembly graph: From a text prompt, the agent decomposes the object into named parts and maps how those parts should connect.
• Solves placement through mates: Parts are joined with typed, machine-checkable mates. Their positions are solved from mated frames rather than guessed directly by the language model.
• Validates incrementally: A part is admitted only after compilation, mate, and connectivity checks succeed.
• Uses a separate visual critic: A decoupled vision module diagnoses visible problems and applies one targeted fix at a time.
• Preserves more than shape: The graph can also carry per-part materials and articulation that has been validated in simulation.

Why this matters

The important shift is in the output representation. Conventional 3D generation tends to produce an end result whose internal construction is difficult to recover. Procedura aims to produce the construction itself: a program describing parts, dimensions, placements, and connections. That representation is much closer to the needs of CAD workflows, robotics simulation, interactive assets, and manufacturing-oriented design.

The framework also creates a useful division of labor. The language model handles specification, decomposition, and code writing. Deterministic or geometric tools check whether the resulting program compiles and whether its assembly relationships make sense. A vision critic then addresses appearance-level issues. Combining these roles can be more robust than relying on image similarity alone when the target has precise topology and mechanical structure.

The paper evaluates Procedura with the assembly judge in P3D-Bench and applies the same judge to MechBench-36, a hard-surface benchmark introduced for this setting. Its abstract reports that the method outperforms native 3D generators and previous 3D-code agents on both evaluations, while producing the sharpest edges among the tested methods. The more distinctive result is representational: Procedura is described as the only evaluated approach that outputs an editable, part-structured program.

The approach does not remove every challenge. Reliable decomposition of complex objects, the design of constraints, visual diagnosis, and compatibility with practical modeling kernels will all affect deployment. Still, it points toward a broader role for 3D agents. The goal may no longer be merely to draw a plausible object, but to organize, verify, and maintain a design that people can continue to modify.

Source: Hugging Face Daily Papers