Arbor introduces constraint meshes for controllable 3D asset generation

Text and image conditioned 3D models often lack direct control over spatial constraints, such as fitting a chair into a specific seating envelope or ensuring clearance for motion. Arbor addresses this by introducing constraint meshes as a native 3D control interface for text-conditioned latent 3D generation.

How Arbor Works

Arbor operates as a trainable attachment within a frozen denoiser, converting constraint meshes into tokens to maintain geometric signals. The interface defines three specific types of regions:

• Hull regions: Areas where geometry should exist.
• Avoidance regions: Areas that must remain empty.
• Touch regions: Areas the object should contact.

Unlike traditional completion or scaffold controls, these meshes act as local typed requirements rather than target evidence, allowing for precise spatial intent without requiring surface data in every region.

Performance and Evaluation

The system was evaluated on automatic and artist-curated benchmarks featuring hull, avoidance, and touch constraints. Results indicate that Arbor improves constraint obedience while preserving object quality and variation, even without dedicated compliance losses. Metric trends from these benchmarks align with user preference studies, confirming the effectiveness of the explicit geometric conditioning.