Source-linked AI summary

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

arXiv:2509.08643v2cs.GRcs.CV

TL;DR

Existing part-based 3D generation methods provide limited controllability and may fail to produce semantically meaningful decompositions, despite the importance of parts for downstream tasks. X-Part addresses this with bounding-box prompts, semantic point-feature injection, synchronized diffusion, and interactive editing, achieving state-of-the-art part-level decomposition and generation performance. The method also supports downstream workflows such as UV unwrapping while retaining editable part control.

  • Problem

    Existing approaches struggle to decompose complete 3D shapes into semantically meaningful parts while recovering geometrically plausible structures for downstream tasks.

  • Method

    X-Part uses bounding-box prompts, point-wise semantic features, synchronized multi-part diffusion, and a training-free editing pipeline for controllable part generation.

  • Results

    X-Part achieves state-of-the-art performance in part-level decomposition and generation across extensive experiments.

  • Takeaways & Limitations

    X-Part supports interactive part editing and can simplify UV unwrapping by producing compact, semantically meaningful part-based UV maps.

  • Takeaways & Limitations

    The method relies on geometric cues without physical-principle guidance, and simultaneous processing makes inference slower as the number of parts increases.

Abstract

from arXiv · show

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.

1 Tencent Hunyuan, 2ShanghaiTech, 3NJU, 4HKU, 5ZJU, 6CUHK

The paper identifies equal contributors, a project leader, and a corresponding author.

  • The paper marks some authors as equal contributors.
  • The paper identifies a project leader.
  • The paper identifies a corresponding author.

1 Introduction

X-Part addresses the difficulty of generating semantically meaningful and geometrically plausible 3D parts for practical creation pipelines. It uses bounding-box prompts, semantic point features, and synchronized diffusion to produce controllable part decompositions, achieving state-of-the-art performance in reported experiments.

  • Existing generative approaches mainly produce monolithic models, while part generation must preserve semantic meaning and plausible internal geometry.
  • X-Part uses P3-SAM to generate initial segmentations, bounding boxes, and semantic features before synchronized multi-part diffusion.
  • Bounding boxes provide coarse location and scale prompts, while point-wise semantic features supply more robust semantic representations.
  • X-Part achieves state-of-the-art performance in part-level decomposition and generation across extensive experiments.
  • The framework includes an editable part-generation pipeline supporting multiple interactive editing methods.

2 Related Work

Prior work spans supervised and foundation-model-based segmentation, object-level 3D generation, and emerging part-level generation methods. X-Part is presented within this landscape as a diffusion architecture for guided decomposition.

  • Earlier part segmentation methods rely on extensive annotations, generalize poorly beyond seen categories, and have limited semantic scalability.
  • Object-level 3D generation methods extend diffusion-based image-generation ideas but can suffer from slow optimization and geometric inconsistency.
  • Part-level generation methods include multi-view segmentation, physical regularization, and DiT-based generative approaches.
  • X-Part’s architecture extracts point features and part bounding boxes, forms global and part conditions, and injects them into multi-part diffusion.

3 Method

X-Part uses object and part-level conditions in a synchronized latent diffusion framework to generate semantically meaningful, structurally coherent 3D parts. Bounding-box cues, semantic features, inter-part attention, and an editing pipeline provide controllable generation and manipulation.

  • 3.2.1 Semantic-Aware Shape Conditioning: Bounding-box prompts specify part locations and scales, while perturbation augmentation improves robustness during inference.Points within each box are encoded into the part-level condition, and random translations and scale changes are applied during training.
  • 3.2.1 Semantic-Aware Shape Conditioning: Point-wise P3-SAM semantic features are concatenated with VAE shape tokens to improve decomposition accuracy despite segmentation inaccuracies.Features are interpolated using down-sampled XYZ positions and randomly masked to improve robustness to high-dimensional semantic inputs.
  • 3.2 Framework: X-Part encodes object and part conditions, then simultaneously generates latent tokens for all parts in a synchronized diffusion process.The object condition comes from the VAE encoder, while part conditions are extracted from points inside bounding boxes and enhanced with semantic features.
  • 3.2.2 Synchronized Part Generation: Inter-part attention extends receptive fields across part tokens, improving boundary awareness and structural consistency while cross-attention preserves input geometry.Half of the self-attention layers attend across all parts; cross-attention injects geometric conditions into the generation process.
  • 3.2.2 Synchronized Part Generation: A flow-matching objective trains the DiT network to transport noisy part tokens toward target latent tokens, using fewer tokens per part to accelerate training and inference.The model predicts a velocity field conditioned on object- and part-level features, and reduced per-part token counts exploit lower individual-part complexity.
  • 3.3 Part Editing: The training-free editing pipeline supports split and adjust operations by resampling and denoising only the latent tokens associated with edited bounding boxes.Tokens for unchanged parts remain fixed during interactive generation.

4 Experiments

X-Part is evaluated against existing decomposition and image-to-3D part-generation methods using geometric metrics, qualitative comparisons, applications, and ablations. Results indicate higher geometric quality, refined semantic decomposition, interactive editing, improved UV unwrapping, and contributions from the framework’s components.

  • Implementation Details: Training uses nearly 2.3 million objects with ground-truth part segmentation, remeshed into watertight object and part meshes.The large-scale dataset is reported to improve the generalizability of the diffusion-based decomposition method.
  • Comparison: The evaluation uses 200 ObjaversePart-Tiny samples with Chamfer Distance and F-Score at thresholds 0.1 and 0.5.Objects are normalized before metric computation and evaluated in a pose-agnostic setting.
  • Comparison: X-Part achieves higher geometric quality and better alignment to ground truth, often producing more numerous semantically reasonable parts.The comparison evaluates both overall object geometry and decomposed parts against existing methods.
  • Applications: Bounding-box edits control part shape and coverage, while merging or splitting boxes enables fusion or finer decomposition.Users can adjust bounding-box position and scale to influence generated part geometry.
  • Applications: Part decomposition simplifies UV unwrapping and produces more compact, semantically meaningful UV maps than holistic-mesh unwrapping.The comparison unwraps holistic meshes versus each separated part independently.
  • Ablation Study: Ablations attribute performance improvements to attention mechanisms, part embeddings, object- and part-level conditions, and semantic point features.The components respectively support contextual part representations, part distinctiveness, geometric priors, location and scale information, and coherent decomposition.

5 Conclusion and Limitation

X-Part decomposes complete 3D objects into structured parts using bounding-box controls and semantic point features, supporting editable generation and downstream content-creation workflows. Its main limitations are the lack of physical-principle guidance and increasing inference time as part count grows.

  • Conclusion: X-Part is a geometry-based framework that uses bounding boxes to generate structured parts from complete 3D objects.Semantic point features improve structural understanding, while users can alter decomposition through bounding-box adjustments.
  • Conclusion: The editable pipeline supports intuitive decomposition control and can generate up to 50 distinct parts for practical applications.The method is intended to reduce downstream 3D content-creation complexity.
  • Limitation: Decomposition lacks physical-principle guidance, which may limit applications requiring specific physically informed decompositions.This is identified as a current limitation of the geometry-cue-based approach.
  • Limitation: Inference time increases with the number of simultaneously processed parts, challenging real-time use for high-part-count objects.All part latent codes are processed together through the diffusion model.
Loading 2509.08643v2…