PaperlayerView exact revision on arXiv

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 3D generation methods often produce monolithic models, while part-level generation must preserve semantic meaning and geometric plausibility. X-Part uses bounding-box prompts and semantic point features in a diffusion framework to decompose objects into parts. It achieves state-of-the-art performance in part-level decomposition and generation.

  • Problem

    Most existing generative approaches produce monolithic 3D models, while part-level generation must preserve meaningful semantics and geometrically plausible structures for downstream tasks.

    Source: §1 · p. 2

  • Method

    X-Part uses a diffusion framework with bounding-box prompts and semantic point features to decompose holistic 3D objects into semantically meaningful, structurally coherent parts.

    Source: §1 · p. 2, §5 · p. 11

  • Results

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

    Source: §1 · p. 2

  • Takeaways & Limitations

    X-Part offers intuitive control and flexible part-level editing, with easier integration into 3D content-creation pipelines.

    Source: §4.3 · p. 10, §5 · p. 11

    X-Part relies on geometric cues without physical-principle guidance, which may limit application-specific decomposition.

    Source: §5 · p. 11

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 Introduction

X-Part is a controllable diffusion framework that decomposes holistic 3D meshes into semantically meaningful, structurally coherent parts, using bounding-box prompts and point-wise semantic features and supporting interactive editing.

1 Introduction

  • Part-level generation must preserve semantic meaning and plausible internal geometry, while existing methods either depend on error-prone segmentation or lack controllability and produce ambiguous boundaries.Meaningful decomposition can also facilitate downstream mesh retopology and UV unwrapping.Source: §1 · p. 2
  • X-Part uses P3-SAM to generate initial segmentations, bounding boxes, and semantic features, then decomposes the mesh through synchronized multi-part diffusion.P3-SAM is the state-of-the-art segmenter used to prepare the initial inputs.Source: §1 · p. 2
  • Bounding boxes guide part locations and scales, reducing overfitting to segmentation masks while providing volume-scale information for partially visible parts.They provide coarser guidance than fine-grained, point-level segmentation cues.Source: §1 · p. 2
  • X-Part injects high-dimensional point-wise semantic features into diffusion to guide decomposition with representations less affected by mask-prediction compression.The paper argues these features remain robust despite inaccuracies in segmentation results.Source: §1 · p. 2
  • The editable pipeline supports interactive local edits, including splitting a part into several parts and adjusting part scales.It is based on bounding boxes and follows Lugmayr et al. (2023).Source: §1 · p. 2
  • X-Part achieves state-of-the-art performance in part-level decomposition and generation.The paper reports extensive experiments across various benchmarks.Source: §1 · p. 2

2 Related Work

Related work spans 3D part segmentation, object-level shape generation, and part-level generation. These approaches expand semantic and generative capabilities, but reported limitations include failures on occluded geometry, limited controllability, and reduced geometric quality.

Part Segmentation

Object-level Shape Generation

Part-level Shape Generation

3 Method

X-Part is a diffusion-based method for controllably decomposing an object point cloud into parts. It combines object- and part-level cues for synchronized generation and provides an editable pipeline for interactive part generation.

3.1 Preliminary

  • The preliminary framework encodes sampled mesh points into VAE latent tokens and decodes spatial queries to signed distance values; the VAE is fine-tuned on part-level data.The encoder uses cross-attention followed by self-attention layers, and the decoder uses self-attention followed by a final cross-attention module.Source: §3.1 · p. 4

3.2.1 Semantic-Aware Shape Conditioning

  • Bounding-box cues and point-wise semantics condition X-Part’s generation of parts from an object point cloud.The object-level condition comes from the VAE-encoded input; points inside each part’s box form a learned part-level condition, and P3-SAM semantics are aligned with shape tokens and randomly masked. Box translation and scaling augmentations improve robustness, while semantic features and inter-part attention help exclude points from adjacent parts. Semantic features use P3-SAM Ma et al. (2025).Source: §3.2.1 · p. 5

3.2.2 Synchronized Part Generation

  • X-Part jointly generates all parts’ latent tokens with DiT blocks, combining intra-part and expanded inter-part attention, geometric cross-attention, and unique learnable part embeddings.The inter-part attention extends the receptive field in half of the self-attention layers to improve awareness across part boundaries; cross-attention supports structural consistency and preserves geometric details.Source: §3.2.2 · pp. 5–6
  • X-Part trains with flow matching, predicting the velocity that transports Gaussian-noised part tokens toward target latents under object- and part-level conditions.Using fewer tokens per part during VAE fine-tuning and X-Part training accelerates training and inference while retaining performance.Source: §3.2.2 · p. 6

3.3 Part Editing

  • A training-free pipeline supports split and adjust edits by resampling and denoising tokens for indicated parts while keeping other parts’ tokens unchanged.Split divides a bounding box to generate several parts; adjust changes a box and regenerates that part and adjacent parts, following Repaint Lugmayr et al. (2022).Source: §3.3 · p. 6

4 Experiments

X-Part is evaluated for part decomposition and holistic shape generation, with comparisons showing stronger decomposition quality and geometric fidelity than the baselines described. The experiments also demonstrate bounding-box-controlled editing and part-aware UV unwrapping, while ablations examine the contributions of the model’s components.

4.1 Implementation Details

  • The implementation uses a 21-block DiT with inter-part attention, part embeddings, and an MoE in the first six blocks.Training uses nearly 2.3 million objects from the P3-SAM dataset Ma et al. (2025), remeshed into watertight object and part meshes; it includes semantic-feature and condition dropout plus bounding-box translation and scaling augmentation.Source: §4.1 · pp. 6–7

4.2 Comparison

  • X-Part outperforms all decomposition baselines, including OmniPart with ground-truth 2D masks, while segmentation-based methods fail to produce complete part geometries.The qualitative comparison also shows higher geometric quality, with more refined decompositions that often contain more semantically reasonable parts.Source: Table 1 · p. 7, §4.2 · pp. 7, 10, Figure 2 · p. 8
  • X-Part produces final objects with higher geometric quality and better ground-truth alignment than the compared image-to-3D part-generation methods.Because part structures may differ across methods, this comparison evaluates the overall geometry composed of all generated parts.Source: Table 2 · p. 9, §4.2 · pp. 9–10

4.3 Applications

  • Bounding-box edits control part shape and coverage; merging boxes fuses parts, while splitting a box creates finer structures.The application also demonstrates that unwrapping parts separately simplifies UV unwrapping and produces more compact, semantically meaningful UV maps.Source: §4.3 · p. 10, Figure 4 · p. 10

4.4 Ablation Study

  • Ablations attribute performance contributions to attention, part embeddings, object- and part-level conditions, and semantic point features.Attention supports part-level latents with global context; object conditioning supplies overall shape priors, part conditioning supplies coarse location and scale, and semantic features support structurally coherent decomposition.Source: §4.4 · pp. 10–11, Figure 5 · p. 11

5 Conclusion and Limitation

X-Part is a geometry-based framework for generating structured 3D parts from bounding boxes, using semantic point features and editable decomposition. It improves geometric quality and fidelity but lacks physical-principle guidance, and inference slows as part counts rise.

Conclusion

  • X-Part takes bounding boxes as input to decompose complete 3D objects into structured parts, letting users control decomposition strategies through box adjustments.The approach supports generation of up to 50 distinct parts.Source: §5 · p. 11
  • Semantic point features provide high-level shape semantics, while the method preserves geometric quality and fidelity and integrates more easily into 3D content-creation pipelines.The paper says this easier integration reduces the complexity of downstream tasks.Source: §5 · p. 11

Limitation

  • The method relies on geometric cues rather than physical principles, which may limit its suitability for some application-specific decomposition requirements.Source: §5 · p. 11
  • Inference time increases with the number of parts because the diffusion model processes all part latent codes simultaneously, challenging real-time use for high-part-count objects.Source: §5 · p. 11

Related papers

Loading 2509.08643v2…