CubePart: An Open-Vocabulary Part-Controllable 3D Generator

TL;DR

CubePart: An open-vocabulary, part-controllable 3D generator for semantic structure generation.

cs.AI 🔴 Advanced 2026-05-28 42 views
Yiheng Zhu Kangle Deng Jean-Philippe Fauconnier Inaki Navarro Daiqing Li Ava Pun Yinan Zhang Peiye Zhuang Xiaoxia Sun Maneesh Agrawala Kiran Bhat Tinghui Zhou
3D generation part control open vocabulary game engines AI

Key Findings

Methodology

CubePart employs a two-stage generative architecture, first performing global shape synthesis and then part-level decoding. It constructs a large open-vocabulary, part-labeled 3D dataset using a scalable data pipeline. Vision-language models and a novel 3D-aware annotation strategy ensure the generated 3D models adhere to user-defined semantic structures.

Key Results

  • Result 1: On a dataset of 462K assets and about 2M parts, CubePart generates 3D models that can be directly used in game engines without manual post-processing.
  • Result 2: Compared to existing methods, CubePart excels in part control and semantic structure consistency, significantly improving generation quality.
  • Result 3: Ablation studies confirm the importance of cross-part attention mechanisms in maintaining global geometric coherence.

Significance

CubePart holds significant value for academia and industry. It addresses the issue of existing 3D generative models failing to meet specific application needs, providing directly usable 3D assets for games and simulations, reducing the need for manual post-processing.

Technical Contribution

CubePart's technical contributions include introducing an open-vocabulary part control mechanism, offering higher semantic control precision and flexibility compared to existing methods. Its two-stage generative architecture and cross-part attention mechanism open new engineering possibilities in 3D generation.

Novelty

CubePart is the first to achieve open-vocabulary 3D part control generation, providing user-defined semantic structure generation capabilities compared to existing fixed vocabulary or implicit inference methods.

Limitations

  • Limitation 1: Inconsistent generation may occur when handling complex semantic structures.
  • Limitation 2: Dependence on large-scale datasets may limit performance on smaller datasets.
  • Limitation 3: Generated parts may require additional adjustment in specific application scenarios.

Future Work

Future work includes optimizing the generative model to improve consistency in complex semantic structures and exploring performance improvements on smaller datasets.

AI Executive Summary

In modern games and simulations, 3D assets often need to be decomposed into specific semantic parts to support animation, physics, and scripted behaviors. However, most generative 3D models either produce monolithic meshes or arbitrary part decompositions that cannot align with application-specific requirements. CubePart presents a generative framework for open-vocabulary, part-controllable 3D mesh generation, allowing users to explicitly control part structure at inference time. By introducing a scalable data pipeline and a two-stage generative architecture, CubePart generates 3D models that adhere to user-defined semantic structures. These assets can be directly integrated into game engines and driven by animation and behavior scripts without manual post-processing. CubePart's innovation lies in its open-vocabulary part control mechanism, offering new engineering possibilities in 3D generation. While CubePart still has room for improvement in generating consistency for complex semantic structures, its impact on academia and industry is undeniable.

Deep Analysis

Background

In modern games and simulations, 3D assets often need to be decomposed into specific semantic parts to support animation, physics, and scripted behaviors. However, most generative 3D models either produce monolithic meshes or arbitrary part decompositions that cannot align with application-specific requirements. CubePart presents a generative framework for open-vocabulary, part-controllable 3D mesh generation, allowing users to explicitly control part structure at inference time.

Core Problem

Existing 3D generative models struggle to generate semantic structures that meet specific application needs. Most models either produce monolithic meshes or arbitrary part decompositions, leading to significant manual post-processing in games and simulations.

Innovation

CubePart's innovation lies in its open-vocabulary part control mechanism. By introducing a scalable data pipeline and a two-stage generative architecture, CubePart generates 3D models that adhere to user-defined semantic structures. These assets can be directly integrated into game engines and driven by animation and behavior scripts without manual post-processing.

Methodology

  • �� Data Pipeline: Constructs a large open-vocabulary, part-labeled 3D dataset.
  • �� Global Shape Synthesis: Generates a complete 3D mesh.
  • �� Part-Level Decoding: Decomposes the complete mesh into parts adhering to user-defined semantic structures.
  • �� Cross-Part Attention Mechanism: Ensures global geometric coherence.

Experiments

The experimental design includes testing on a dataset of 462K assets and about 2M parts. Ablation studies confirm the importance of cross-part attention mechanisms in maintaining global geometric coherence. Compared to existing methods, CubePart excels in part control and semantic structure consistency.

Results

CubePart significantly outperforms existing methods in generation quality and semantic control precision. Experimental results show that CubePart generates 3D models that can be directly used in game engines without manual post-processing.

Applications

CubePart-generated 3D models can be directly used in games and simulations, supporting animation, physics, and scripted behaviors. This reduces the need for manual post-processing and improves development efficiency.

Limitations & Outlook

CubePart may produce inconsistent results when handling complex semantic structures. Additionally, its dependence on large-scale datasets may limit performance on smaller datasets. Future work includes optimizing the generative model to improve consistency in complex semantic structures.

Plain Language Accessible to non-experts

Imagine you're building a Lego model. CubePart is like a smart assistant that, once you tell it what model and parts you want, can find all the right Lego pieces and assemble them into a complete model. You can freely choose the names and number of parts, and CubePart ensures each part meets your requirements, like a perfect puzzle. Even if you want a complex model, like a turtle with a castle, CubePart can handle it effortlessly.

ELI14 Explained like you're 14

Imagine you're playing a super cool game and want a unique race car. CubePart is like a magic tool that, once you tell it what you want your car to look like, can generate a perfect 3D model for you. You can choose parts like the body, wheels, and engine, and CubePart ensures each part fits your requirements, like a perfect puzzle. Isn't that amazing?

Glossary

3D Generation

The process of creating three-dimensional models using computer algorithms.

CubePart is used to generate 3D models adhering to user-defined semantic structures.

Part Control

The ability to independently control different parts of a model during generation.

CubePart allows users to explicitly control part structure at inference time.

Open Vocabulary

A set of terms that is not restricted, allowing user-defined entries.

CubePart supports open-vocabulary part control mechanisms.

Cross-Part Attention

A mechanism ensuring geometric coherence among different parts.

CubePart uses cross-part attention mechanisms to maintain global coherence.

Data Pipeline

A process for constructing and processing large-scale datasets.

CubePart introduces a scalable data pipeline for building 3D datasets.

Open Questions Unanswered questions from this research

  • 1 How can CubePart's generation quality be improved on smaller datasets?
  • 2 How can CubePart's consistency in generating complex semantic structures be further optimized?
  • 3 How can CubePart's part control mechanism adapt to diverse needs in different application scenarios?

Applications

Immediate Applications

Game Development

Game developers can use CubePart to generate 3D models that meet specific needs, improving development efficiency.

Simulation Training

CubePart-generated 3D models can be used in simulation training, supporting complex animations and physical behaviors.

Long-term Vision

Virtual Reality

CubePart has the potential to generate dynamic 3D environments in virtual reality, enhancing user experience.

Abstract

Interactive 3D assets used in games and simulation are typically decomposed into specific semantic parts to support animation, physics, and scripted behaviors, yet most generative 3D models produce either monolithic meshes or arbitrary part decompositions that cannot be aligned with application-specific requirements. We present CubePart, a generative framework for open-vocabulary, part-controllable 3D mesh generation that exposes part structure as an explicit inference-time control signal. Given a global text prompt and a user-defined parts schema expressed as an open-ended list of part names, our method generates a set of meshes - one per schema element - that assemble into a coherent object while respecting the specified semantic structure. To enable this capability, we introduce a scalable data pipeline to construct a large open-vocabulary, part-labeled 3D dataset, along with a two-stage generative architecture that separates global shape synthesis from part-level decoding. We demonstrate that the resulting assets can be directly integrated into game engines and driven by animation and behavior scripts without manual post-processing. Project Page: https://cubepart.github.io/

cs.AI