PaperlayerView exact revision on arXiv

Source-linked AI summary

FullPart: Generating each 3D Part at Full Resolution

Lihe Ding, Shaocong Dong, Yaokun Li, Chenjian Gao, Xiao Chen, Rui Han, Yihao Kuang, Hong Zhang, Bo Huang, Zhanpeng Huang, Zibin Wang, Dan Xu, Tianfan Xue

arXiv:2510.26140v1cs.CV

TL;DR

Part-based 3D generation needs detailed, semantically meaningful parts, but prior implicit and shared-grid voxel approaches constrain part detail. FullPart combines implicit layout generation with isolated full-resolution part grids and center-corner encoding; the authors report state-of-the-art results across multiple metrics.

  • Problem

    Implicit part representations can lack geometric detail, while shared global voxel grids can allocate too few voxels to small parts.

    Source: §1 · p. 2

  • Method

    FullPart generates layouts with implicit representations, detailed parts in isolated full-resolution grids, and uses center-corner encoding to support coherence.

    Source: §1 · p. 2

  • Results

    FullPart achieves state-of-the-art performance across multiple part-level 3D generation metrics, particularly for complex part interactions and plausible occluded geometries.

    Source: §1 · p. 3

  • Takeaways & Limitations

    PartVerse-XL provides human-annotated part data at scale to address the scarcity of reliable part-level 3D training data.

    Source: §1 · pp. 2–3

    The pretrained 3D generator's positional encoding layers were trained only at 64 × 64 × 64 resolution, so the method relies on positional-encoding extrapolation during finetuning.

    Source: §3.3 · p. 6

Abstract

from arXiv · show

Part-based 3D generation holds great potential for various applications. Previous part generators that represent parts using implicit vector-set tokens often suffer from insufficient geometric details. Another line of work adopts an explicit voxel representation but shares a global voxel grid among all parts; this often causes small parts to occupy too few voxels, leading to degraded quality. In this paper, we propose FullPart, a novel framework that combines both implicit and explicit paradigms. It first derives the bounding box layout through an implicit box vector-set diffusion process, a task that implicit diffusion handles effectively since box tokens contain little geometric detail. Then, it generates detailed parts, each within its own fixed full-resolution voxel grid. Instead of sharing a global low-resolution space, each part in our method - even small ones - is generated at full resolution, enabling the synthesis of intricate details. We further introduce a center-point encoding strategy to address the misalignment issue when exchanging information between parts of different actual sizes, thereby maintaining global coherence. Moreover, to tackle the scarcity of reliable part data, we present PartVerse-XL, the largest human-annotated 3D part dataset to date with 40K objects and 320K parts. Extensive experiments demonstrate that FullPart achieves state-of-the-art results in 3D part generation. We will release all code, data, and model to benefit future research in 3D part generation.

1 INTRODUCTION

FullPart combines implicit layout generation with explicit full-resolution part generation to address limited detail and coherence in prior approaches. The authors also introduce PartVerse-XL and report state-of-the-art part-generation results.

1 INTRODUCTION

  • Part decomposition supports downstream tasks including texture mapping, animation, physical simulation, and fine-grained editing.The introduction frames independently manipulable semantic parts with maintained global coherence as a useful generation goal.Source: §1 · p. 2
  • FullPart uses implicit representations for bounding-box layouts and isolated full-resolution voxel grids for detailed parts, followed by textured-mesh refinement.The design draws on the authors’ view that implicit representations suit layouts, while each explicit part needs its own full-resolution space.Source: §1 · p. 2
  • Center-corner encoding gives tokens spatial-extent information to address scale misalignment between neighboring parts and support smooth stitching.The authors say this also makes finetuning easier.Source: §1 · p. 2
  • PartVerse-XL contains 40K objects and 320K parts with part-aware texture descriptions, using human refinement to improve annotation quality.The dataset is built from Objaverse-XL objects using mesh pre-segmentation followed by human refinement.Source: §1 · pp. 2–3
  • FullPart achieves state-of-the-art part-level 3D generation across multiple metrics, with particular strength in complex part interactions and plausible occluded geometries.The reported experiments also highlight plausible geometries for small parts.Source: §1 · p. 3

2 RELATED WORK

Prior 3D generation work often lacks explicit part decomposition, while part-aware methods face constraints including segmentation quality and shared representation capacity. FullPart addresses the latter bottleneck by treating each part as an independent full-resolution object.

2.1 3D GENERATION

  • Earlier 3D-native generators produce monolithic shapes, while part-aware methods have ranged from category-specific systems to general-purpose approaches with segmentation or patch-information constraints.Recent examples include contextual part latents, joint layout-and-geometry generation, implicit part tokens, part proposals, bounding-box voxel generation, and bounding-box attention.Source: §2.1 · p. 3, §2.2 · p. 3

2.2 PART GENERATION

  • Part-aware methods sharing a global representation space give small parts limited latent capacity or very low effective voxel resolution; FullPart treats each part as an independent full-resolution object.The paper identifies this shared-space bottleneck as inadequately addressed by prior work.Source: §2.2 · pp. 3–4
  • Figure 2 presents FullPart as three stages: implicit vecset layout generation, explicit full-resolution voxel generation for each part, and coarse-structure refinement into textured meshes.Source: Figure 2 · p. 4

3 METHODOLOGY

FullPart turns an image or text prompt into a structured object by generating part layouts, full-resolution part structures, and refined textured meshes. Its methodology combines implicit and explicit representations while using attention and spatial encoding to coordinate parts.

3 METHODOLOGY

  • The framework generates bounding-box layouts, coarse part structures, and refined textured meshes from an image or text prompt.The stages use implicit vecset diffusion for layouts, explicit per-part voxel grids for structures, and mesh decoding for refinement.Source: §3 · p. 4

3.1 PRELIMINARY: 3D OBJECT GENERATION FRAMEWORKS

  • Implicit vecsets represent geometry as latent tokens decoded to an SDF, while explicit sparse voxels encode positions and local geometry or appearance.Part-aware generation uses intra-part and inter-part attention to model local details and relationships across parts.Source: §3.1 · pp. 4–5

3.2 LAYOUT GENERATION

  • Bounding boxes are represented as cuboid meshes encoded into latent tokens and generated with a global branch and intra- and inter-box attention.The global branch predicts holistic structure while the attention strategy models box-specific characteristics and relationships.Source: §3.2 · p. 5

3.3 3D COARSE STRUCTURE GENERATION

  • Each part occupies its own full-resolution voxel grid, while center-and-corner positional encoding supplies global spatial context across parts of different sizes.This encoding addresses token misalignment caused by voxels representing different spatial extents, while retaining the pretrained model architecture.Source: §3.3 · pp. 5–6

3.4 REFINEMENT

  • The refinement stage generates features for occupied voxels and decodes them into textured part meshes, using independent part rendering during training.At inference, the model requires only a single global image or text input.Source: §3.4 · pp. 6–7

3.5 OPTIMIZATION LOSS

  • All framework stages are trained with a Conditional Flow Matching objective.The passage identifies ϵθ as diffusion noise and vθ as the predicted vector field.Source: §3.5 · p. 7

4 PARTVERSE-XL DATASET

PartVerse-XL provides a large, refined dataset of part-aware 3D objects, pairing textured parts with descriptive captions. Its construction combines automated pre-segmentation with expert refinement.

4 PARTVERSE-XL DATASET

  • PartVerse-XL contains 40K high-quality objects and 320K semantically consistent, textured parts across over 200 categories, each paired with a descriptive caption.The dataset is an expanded and refined extension of PartVerse.Source: §4 · p. 7
  • Experts refine over-segmented outputs by merging or splitting components for semantic clarity, structural symmetry, and texture preservation, while discarding low-quality or ambiguous assets.The pipeline first combines geometric priors with semantic cues from SAM-2 and Samesh to create the over-segmented outputs.Source: §4 · p. 7
  • Captions describe each part’s shape, appearance, material, and relationship to its object using a selected multi-view composite image.The composite shows the part bounding box on the view with maximal part–whole visibility overlap and is input to a vision-language model.Source: §4 · p. 7
  • PartVerse-XL is presented as establishing a new standard in scale, semantic fidelity, and multimodal alignment for part-level 3D generation.The passage points to supplementary material for further dataset details.Source: §4 · p. 7

5 EXPERIMENTS

Experiments compare FullPart with part-aware and full-object generators, evaluate its metrics on a 100-object test set, and examine design choices and editing applications. The reported comparisons favor FullPart in part-level fidelity and global coherence, while ablations illustrate the roles of encoding, annotations, and per-part resolution.

5 EXPERIMENTS

  • Experiments use the PartVerse-XL training set of 40K objects and 320K parts and a dedicated test set of 100 objects.Training uses three sequential stages on eight NVIDIA A100 GPUs; inference clamps part count at 30 and uses isolated 64^3 grids.Source: §5 · p. 8, §5.1 · p. 8

5.2.1 COMPARISON WITH STATE-OF-THE-ART PART GENERATORS

  • Against PartCrafter and Omnipart, FullPart shows stronger geometric fidelity and structural coherence, particularly for intricate assemblies and occluded regions.The passage attributes fragmented parts to PartCrafter’s implicit token entanglement and voxelization artifacts in small components to Omnipart.Source: §5.2.1 · p. 8

5.2.2 COMPARISON WITH FULL-OBJECT METHODS

  • Compared with TRELLIS and Direct3D-S2, FullPart preserves fine-grained details that full-object methods lose in sparse global grids.The passage highlights robotic head features as an example.Source: §5.2.2 · p. 8

5.2.3 QUANTITATIVE EVALUATION

  • FullPart outperforms all baselines in part-level and full-level metrics on the 100-object test set, with Part-CD inapplicable to methods lacking bounding-box-conditioned part generation.The reported metrics are global F-Score, global mesh chamfer distance, Part-CD using the same layout boxes, and ULIP Score; the supplied passages do not state metric values.Source: §5.2.3 · p. 8, Table 1 · p. 9

5.3 ABLATION STUDIES

  • Replacing center-and-corner coordinates with center-only encoding fails to model part interactions, including aligned chair legs.The ablation supports the importance of explicit location and scale information for spatial relationships between parts.Source: §5.3 · pp. 8–9
  • Metadata-only training yields semantically incorrect parts, whereas human-refined annotations yield coherent, functionally meaningful parts.Figure 6 compares metadata-derived structural information with human-annotated data.Source: §5.3 · p. 9, Figure 6 · p. 9
  • A shared global voxel grid degrades detail in small parts, while dedicated per-part grids maintain resolution regardless of relative size.The comparison uses per-part N^3 grids with N = 64 versus a shared global N^3 space.Source: §5.3 · p. 9, Figure 6 · p. 9

5.4 APPLICATION

  • Users can add, delete, or reshape layout boxes to edit corresponding parts; unchanged components are regenerated consistently by injecting their clean latent tokens as fixed inputs.The example adds two accessories to a rifle and elongates its barrel through box insertion and stretching.Source: §5.4 · p. 9

6 CONCLUSION

The conclusion presents FullPart as combining implicit and explicit representations, with each part generated at full resolution to address shared-grid resolution limits. It also introduces PartVerse-XL as an annotated dataset for future part-generation research.

6 CONCLUSION

  • FullPart integrates implicit and explicit representations and treats each part as a full-resolution object to overcome the resolution bottleneck of shared voxel grids.The conclusion also presents PartVerse-XL as a large-scale annotated dataset for advancing future 3D part-generation research.Source: §6 · p. 9

A.1 NETWORK ARCHITECTURE

The network architecture combines inter-part and intra-part attention while retaining a holistic generation branch and pretrained model components. These design choices preserve global semantic context during fine-tuning.

A.1 NETWORK ARCHITECTURE

  • The three-stage models start from pretrained holistic generators: TripoSG for layout generation and TRELLIS stages for coarse structure and mesh refinement.The layout model replaces original tokens with box tokens and adds box ID embeddings.Source: §A.1 · p. 13
  • Each stage uses inter-part and intra-part attention blocks and retains a holistic generation branch to preserve global semantic context and stabilize fine-tuning.Image or text conditions enter through cross-attention, and all part tokens attend to all condition embeddings.Source: §A.1 · p. 13

A.2 TRAINING AND INFERENCE DETAILS

Training caps each object at 30 parts by selecting the largest parts, while inference can generate more through sequential sampling; the passage also specifies conditional-generation settings.

A.2 TRAINING AND INFERENCE DETAILS

  • Training uses at most 30 parts per object, selecting the 30 largest by bounding-box size.The cap is due to GPU memory constraints.Source: §A.2 · p. 13
  • Inference supports objects with more than 30 parts by sampling 30 parts at a time while retaining the previous denoised global token to maintain global structure.The initial round includes a global part; later rounds replace global tokens with noisy versions of the previous denoised global token at each timestep.Source: §A.2 · p. 13
  • Training renders 24 conditional images per object and randomly selects one, with condition tokens dropped at probability 0.1.Inference applies classifier-free guidance with scale 3.5.Source: §A.2 · p. 13
  • Bounding-box augmentation during coarse-structure training is used to make the model robust to imperfect box inputs.Source: §A.2 · p. 13

B PARTVERSE-XL DATASET

PartVerse-XL combines large-scale, human-refined part annotations with part-level text descriptions, targeting semantic consistency and texture fidelity for part-aware generation.

B PARTVERSE-XL DATASET

  • PartVerse-XL contains 40K Objaverse-XL objects and 320K semantically consistent, texture-preserving parts across more than 200 object categories, each with a detailed part-level textual description.The passage says its scale and diversity surpass existing part-level benchmarks such as PartNet Mo et al. (2018).Source: B PARTVERSE-XL DATASET · p. 13
  • Dataset construction combines automated over-segmentation using 3D cues and semantic priors with human refinement of part boundaries.Annotators merge fragments or split conflated regions, while preserving textures and materials.Source: B PARTVERSE-XL DATASET · pp. 13, 15
  • For each part, captions are generated from a composite image pairing a whole-object render with an isolated-part view and a bounding box around the part.A vision-language model produces descriptions of shape, appearance, material, and functional or spatial relationships to the parent object.Source: B PARTVERSE-XL DATASET · p. 15
  • PartVerse-XL is described as an order-of-magnitude larger than prior part datasets, with improved semantic consistency, texture fidelity, and multimodal alignment.Source: B PARTVERSE-XL DATASET · p. 15
Loading 2510.26140v1…