✦
Summary
Arbor 3D is a controllable 3D asset generation method that uses explicit geometric conditioning. It introduces constraint meshes as a native control interface for telling a text-conditioned 3D generator where geometry should exist, where it should stay empty, and which surfaces should be touched.
Key Features
Adds explicit geometry constraints to text-conditioned 3D generation.
Uses hull regions for where object geometry should exist.
Uses avoidance regions for space that should remain empty.
Uses touch regions for surfaces the generated object should contact.
Builds on frozen TRELLIS.2-style generators and geometry encoders.
Routes local constraint memory into sparse 3D denoising blocks.
Supports variation under fixed constraints instead of collapsing to one shape.
Provides preprint links, code/model references, and direct demonstration videos.
About Arbor 3D
The method keeps the text-conditioned generator and geometry encoders frozen while training geometry projection, positional embedding, summary modules, and grounding adapters. Hull, avoidance, and touch meshes are encoded into compact geometry memory and routed to local regions of the TRELLIS sparse structure grid.
Arbor 3D is useful for artists and 3D generation researchers who need direct spatial control rather than only text or image prompts. It can preserve plausible object generation while satisfying typed geometric requirements and allowing variation in unconstrained regions.