Treellis: Exposing Editable Assembly Structures for Progressive Human-AI Co-Creation of 3D Models

Overview

Generative AI lowers the barriers to 3D modeling but offers limited control over outcomes. Prompt-only tools return a finished model with little visibility into how it is constructed, so fixing one part often means re-prompting and watching unrelated parts change.

Treellis is a human-AI co-creation tool for 3D modeling. Given a text prompt, it produces a 3D object along with an interface that reveals its component-level assembly structure and in-situ editable parameters. Users can inspect and iteratively refine problematic regions at multiple levels of granularity, using both AI-assisted operations and direct manual edits, while keeping the parts that already work.

This work received a Best Paper Award at ACM UIST 2026.

How It Works

Treellis keeps one shared scene representation behind four synchronised views: the Prompt Input, the 3D Viewport, the Structure Tree and the Parameter Panel.

Generate: Describe an object in natural language. Treellis returns a 3D model together with its assembly structure and object-specific parameters.
Inspect & Target: Hover or click in the Structure Tree or the 3D viewport. Selections are cross-highlighted, so users always know which part they are about to change.
Refine Locally: Adjust parameters with sliders, add, remove or reorganise nodes by hand, or prompt the AI on just the selected subtree, leaving the rest of the model untouched.

Structure Tree

The Structure Tree shows how the object is assembled from parts, operations and repeats. It makes the otherwise hidden construction of an AI-generated model visible and directly editable.

Parameter Panel

The Parameter Panel exposes part-level sliders for position, rotation, scale, size and colour, so small corrections do not require another round of prompting.

Evaluation

We evaluated Treellis in two studies:

A controlled between-subjects study (N=24) against a prompt-only baseline. Treellis raised independent task completion from 92/108 to 106/108 and improved perceived usability (SUS 87.7 vs. 74.0).
A multi-session open-ended design study (N=5), in which participants developed efficient hybrid strategies that mix AI generation with manual edits to build complex designs, from toy cars and a bicycle to a courtyard and a heat exchanger with internal structure.