X-Part: high fidelity and structure coherent shape decomposition
Xinhao Yan, Jiachen Xu, Yang Li, Changfeng Ma, Yunhan Yang, Chunshi Wang, Zibo Zhao, Zeqiang Lai, Yunfei Zhao, Zhuo Chen, Chunchao Guo
2025-09-15
Summary
This paper introduces a new system called X-Part that focuses on creating 3D models by building them from individual, recognizable parts, like assembling LEGOs.
What's the problem?
Currently, creating 3D models piece-by-piece is difficult because existing methods don't give users enough control over the process and often break down complex objects into parts that don't make logical sense or have messy connections. It's hard to get a good, clean separation of a 3D object into its meaningful components.
What's the solution?
X-Part solves this by using bounding boxes – essentially outlining where each part should be – as a starting point for generating the individual pieces. It also uses information about what each part *is* to make sure the decomposition is logical. The system is designed to be interactive, allowing users to edit and refine the parts as they're created. It essentially learns to build 3D objects from their component parts based on simple prompts and semantic understanding.
Why it matters?
This work is important because it makes it easier to create 3D models that are ready for things like 3D printing, adding textures, or being used in video games. By creating models from well-defined parts, it’s easier to modify and work with them, leading to a new way to build high-quality, editable 3D assets.
Abstract
Generating 3D shapes at part level is pivotal for downstream applications such as mesh retopology, UV mapping, and 3D printing. However, existing part-based generation methods often lack sufficient controllability and suffer from poor semantically meaningful decomposition. To this end, we introduce X-Part, a controllable generative model designed to decompose a holistic 3D object into semantically meaningful and structurally coherent parts with high geometric fidelity. X-Part exploits the bounding box as prompts for the part generation and injects point-wise semantic features for meaningful decomposition. Furthermore, we design an editable pipeline for interactive part generation. Extensive experimental results show that X-Part achieves state-of-the-art performance in part-level shape generation. This work establishes a new paradigm for creating production-ready, editable, and structurally sound 3D assets. Codes will be released for public research.