Generating Human-AI Collaborative Design Sequence for 3D Assets via Differentiable Operation Graph

TL;DR

This paper introduces a differentiable operation graph framework for generating human-AI collaborative design sequences in 3D modeling, aligning AI outputs with human design workflows for improved practicality.

Contribution

It proposes a novel differentiable operation formulation and hierarchical graph model to generate design sequences without supervision, enhancing AI-human collaboration in 3D asset creation.

Findings

High geometric fidelity of generated sequences

Smooth mesh wiring and rational step composition

Flexible editing capacity and industry compatibility

Abstract

The emergence of 3D artificial intelligence-generated content (3D-AIGC) has enabled rapid synthesis of intricate geometries. However, a fundamental disconnect persists between AI-generated content and human-centric design paradigms, rooted in representational incompatibilities: conventional AI frameworks predominantly manipulate meshes or neural representations (\emph{e.g.}, NeRF, Gaussian Splatting), while designers operate within parametric modeling tools. This disconnection diminishes the practical value of AI for 3D industry, undermining the efficiency of human-AI collaboration. To resolve this disparity, we focus on generating design operation sequences, which are structured modeling histories that comprehensively capture the step-by-step construction process of 3D assets and align with designers' typical workflows in modern 3D software. We first reformulate fundamental modeling operations (\emph{e.g.}, \emph{Extrude}, \emph{Boolean}) into differentiable units, enabling joint optimization of continuous (\emph{e.g.}, \emph{Extrude} height) and discrete (\emph{e.g.}, \emph{Boolean} type) parameters via gradient-based learning. Based on these differentiable operations, a hierarchical graph with gating mechanism is constructed and optimized end-to-end by minimizing Chamfer Distance to target geometries. Multi-stage sequence length constraint and domain rule penalties enable unsupervised learning of compact design sequences without ground-truth sequence supervision. Extensive validation demonstrates that the generated operation sequences achieve high geometric fidelity, smooth mesh wiring, rational step composition and flexible editing capacity, with full compatibility within design industry.