# Arbor: Stability AI and University of Tübingen Propose Geometric Constraint Meshes for Controllable 3D Generation

_Research · published 2026-06-25_

Researchers from Stability AI and the University of Tübingen have introduced Arbor, a method for controllable 3D asset generation that uses typed geometric constraint meshes as a native control interface. Rather than relying on target shape evidence, Arbor lets artists define hull regions (where geometry should exist), avoidance regions (where space must remain empty), and touch regions (where the object must make contact) — spatial intent that is typically known before generation begins.

Arbor integrates these constraints into the frozen TRELLIS generator by fusing constraint meshes into a geometry memory and routing relevant tokens to each local region of the sparse structure grid via learned adapter modules. Only the geometry projection, positional embeddings, and grounding adapters are trained, leaving the base generator intact. In evaluations on Toys4K benchmarks, Arbor outperformed baselines including TRELLIS, Spice-E, and SpaceControl, winning 59.2% of pairwise preferences in a 27-participant, 404-trial user study.

## Sources
- [arbor.jdihlmann.com](https://arbor.jdihlmann.com/)
---
Canonical: https://genbuzz.news/posts/arbor-stability-ai-and-university-of-tubingen-propose-geometric-constraint-meshes-for-controllable-3d-generation