Stability AI's Arbor gives 3D generation a spatial blueprint

Generating 3D models with AI has long felt like pulling a slot machine handle. You type in a prompt, hit enter, and pray the resulting mesh does not float in midair or break your character rigs. In production, artists usually give up and fix the mesh by hand.

Stability AI just tackled this bottleneck with Arbor, a dedicated control system designed to give 3D generators a rigid spatial blueprint before synthesis even begins.

Hulls, holes, and contact points

Instead of trying to describe spatial dimensions through text, Arbor lets creators draw rough constraint meshes directly in 3D space. These are not finished assets, but simple guide volumes typed into three distinct categories.

The first is hull regions, which mark where geometry must appear, like the cushion of a couch or a tool handle. The second is avoidance regions, which strictly forbid geometry to ensure clearance for mechanical joints or animation movement. The third is touch regions, telling the generator exactly where a surface must connect, such as table legs meeting a flat floor.

For 3D artists, drawing a crude bounding box takes seconds and communicates spatial bounds infinitely better than any text prompt or reference photo could.

A 600MB adapter that leaves the base model intact

Injecting precise geometric conditions into latent diffusion models is notoriously difficult without destroying the generator's creativity. Retraining an entire foundation model from scratch is both expensive and unwieldy.

Arbor solves this by acting as a lightweight, trainable attachment inside a frozen denoiser. Paired with open 3D generators like Trellis and Trellis 2, Arbor converts artist-drawn constraint meshes into tokens and injects them where needed across latent space, leaving the underlying base model untouched.

Because it does not carry the weight of a full generator, the Arbor model is remarkably small at roughly 600MB. Stability AI has made the code and weights publicly accessible on Hugging Face and GitHub.

Why it matters

We see Arbor as a major step toward making AI 3D generation viable in real production pipelines rather than just concept art moodboards.

Until now, generative 3D has been held back by its inability to meet strict spatial envelopes. Game developers reuse animation rigs across hundreds of props. If a weapon grip clips through a hand or a vehicle seat is two inches too high, the asset is unusable. Arbor eliminates that guesswork by anchoring generative freedom inside practical volume constraints.

Architecturally, Stability AI took a smart route here. Building a 600MB pluggable adapter rather than forcing creators onto a proprietary, closed ecosystem makes adoption easy. Anyone already running Trellis locally can plug this in without overhaul.

Keep in mind that Arbor only steers geometry, so overall texture quality and topology still rely on Trellis. But giving creators precise control over occupied and empty space is the exact puzzle piece production pipelines were waiting for.