Part-level 3D generation is crucial for various downstream applications, including gaming, film production, and industrial design. However, decomposing a 3D shape into geometrically plausible and meaningful components remains a significant challenge. Previous part-based generation methods often struggle to produce well-constructed parts, exhibiting poor structural coherence, geometric implausibility, inaccuracy, or inefficiency. To address these challenges, we introduce EI-Part, a novel framework specifically designed to generate high-quality 3D shapes with components, characterized by strong structural coherence, geometric plausibility, geometric fidelity, and generation efficiency. We propose utilizing distinct representations at different stages: an Explode state for part completion and an Implode state for geometry refinement. This strategy fully leverages spatial resolution, enabling flexible part completion and fine geometric detail generation. To maintain structural coherence between parts, a self-attention mechanism is incorporated in both exploded and imploded states, facilitating effective information perception and feature fusion among components during generation. Extensive experiments on multiple benchmarks demonstrate that EI-Part efficiently produces semantically meaningful and structurally coherent parts with fine-grained geometric details, achieving state-of-the-art performance in part-level 3D generation. Project page: https://cvhadessun.github.io/EI-Part/
Figures & tables
Figure 2: The pipeline of our proposed EI-Part. Given an input 3D shape O , we first obtain its normal map {ni}i=16 and canonical coordinate maps (CCMs) {ci}i=16 from six views. We then perform frontal segmentation using SAM and employ MVSegNet to achieve multi-view consistent segmentations. These segmentations are lifted to 3D to create an initial part segmentation, which is subsequently inpainted by InSegNet to produce accurate 3D segmented shapes {psk}k=1K . The segmented result is exploded into discrete 3D voxels, enabling us to perform conditional diffusion completion to generate geometrically plausible and complete part structures ec . Next, the exploded complete parts are imploded back to a compact state for conditional diffusion refinement, capturing fine-level details gd . Thanks to the utilization of the exploded-imploded strategy, we fully utilize spatial resolution at different stages, enhancing the accuracy of part completion and surface details. As a result, the 3D shapes gd generated by our method exhibit individual, structurally consistent, and geometrically plausible part shapes with fine-grained geometric details.
Method
Voxel IOU ↑
CD ↓
Voxel F-Score 0.01 ↑
F-Score 0.1 ↑
F-Score 0.05 ↑
F-Score 0.01 ↑
PartPacker [ 42 ]
0.2586
0.1273
0.3768
0.8199
0.6428
0.2435
PartCrafter [ 24 ]
0.0742
0.3474
0.1316
0.4429
0.2801
0.0749
HoloPart [ 50 ]
0.6106
0.0431
0.7374
0.9557
0.9402
0.6400
X-Part [ 48 ]
0.7478
0.0599
0.8413
0.9256
0.9087
0.7923
OmniPart [ 52 ]
0.2861
0.1431
0.4007
0.7911
0.6516
0.2644
Ours
0.7981
0.0194
0.8452
0.9910
0.9742
0.8129
Table 1: Quantitative comparison between our method and state-of-the-art methods in part generation. Our method achieves the best performance across all evaluation metrics, demonstrating superior shape fidelity and geometric accuracy. Note that BANG is a commercial product with no open-source code available, making quantitative comparison infeasible.
Figure 3: Qualitative comparison between ours and the state of the arts. For each row, the input model (a) is followed by the parts generated by (b) BANG [ 56 ] , (c) HoloPart [ 50 ] , (d) X-Part [ 48 ] , (e) OmniPart [ 52 ] and (f) our method. Our method demonstrates proficiency in generating high-quality 3D shapes with well-defined parts, achieving notable performance in structural coherence, geometric plausibility, fidelity, and efficiency. Please refer to our supplementary materials for a qualitative comparison with additional methods (e.g., PartPacker and PartCrafter).
Figure 4: Qualitative comparison between our method and the state of the art, where all methods share the segmentation results of ours. Each row presents the input model (a), followed by (b) our segmentation, and the parts generated by (c) HoloPart [ 50 ] , (d) X-Part [ 48 ] , (e) OmniPart [ 52 ] and (f) our method.
Figure 5: Qualitative comparison between our method and the alternative strategies. The input is the same segmentation.
Figure 6: Qualitative comparison between our segmentation method and the alternative methods.
Figure 7: Qualitative comparison between ours and the state of the arts. For each row, the input model (a) is shown alongside the parts generated by (b) BANG [ 56 ] , (c) HoloPart [ 50 ] , (d) X-Part [ 48 ] , (e) OmniPart [ 52 ] , (f) PartPacker [ 42 ] , (g) PartCrafter [ 24 ] , and (h) our method. Our method demonstrates exceptional proficiency in generating high-quality 3D shapes with well-defined parts, achieving significant improvements in structural coherence, geometric plausibility, fidelity, and efficiency.
Setting
Part-level
Overall-object
Voxel IOU ↑
CD ↓
Voxel F-Score 0.01 ↑
F-Score 0.1 ↑
F-Score 0.05 ↑
F-Score 0.01 ↑
Voxel IOU ↑
CD ↓
Voxel F-Score 0.01 ↑
F-Score 0.1 ↑
F-Score 0.05 ↑
F-Score 0.01 ↑
w/o Explode
0.7647
0.0153
0.8883
0.9975
0.9866
0.7599
0.9310
0.0092
0.9642
1.0000
1.0000
0.9680
w/o Implode
0.7992
0.0168
0.8664
0.9998
0.9937
0.8678
0.8036
0.0171
0.8896
0.9769
0.9615
0.8243
Whole
-
-
-
-
-
-
0.9005
0.0152
0.9091
0.9975
0.9869
0.8691
Table 2: Quantitative comparison across different settings at both the part and overall object levels. Since the part-level ground truth (GT) is derived from the complete results, the absence of the Explode operation results in lower part completion quality, underscoring its significance. Meanwhile, the explode–implode refinement slightly compromises overall mesh accuracy, revealing a trade-off between enhancing part-level details and maintaining global reconstruction precision.
Part-level 3D assets are essential for editing, reassembly, and interaction, yet recovering such structure from a single image remains challenging due to occlusion, ambiguous boundaries, and the need for coherent multi-part reasoning. Existing approaches struggle to achieve both controllable part-level generation and coherent multi-part structure, as part identity and spatial allocation are typically inferred implicitly. We present Seg3DParts, a segmentation-grounded framework for controllable part-level 3D generation from a single image. By treating segmentation as an explicit grounding signal, our method defines part identity during generation, enabling each component to be anchored to a corresponding image region. To ensure coherent assemblies, we introduce structured cross-part interaction that allows components to exchange global context throughout the generative process. As a result, Seg3DParts directly generates well-aligned part meshes in a shared canonical space without post-hoc alignment, supporting flexible and controllable decomposition. We further introduce PartObjectNet, a large-scale dataset with over 200K objects and 1M annotated parts. Experiments demonstrate that Seg3DParts achieves superior geometry quality, cross-part coherence, and part-level controllability over existing methods.
Jiantao Lin, Meixi Chen, Yingjie Xu +5
The Hong Kong University of Science and Technology (Guangzhou) · Knowin AI · The Hong Kong University of Science and Technology
Generating 3D assets with explicit parts requires two properties at once: geometric detail inside each part and structural coherence across the whole object. Recent image-conditioned 3D generation models adopt a structured 3D representation that anchors tokens to explicit spatial positions, capturing both properties, but their output remains a single fused mesh. We ask how a part-decomposed generation model can inherit both properties of these holistic priors, and answer with STRUCTPARTS, a framework whose core technical contribution is a dual-frame coordinate interleaving mechanism: each shape latent is interpreted under both a part-local and an object-global reference frame, and processed through a global-local inter-leaved transformer that simultaneously reuses the prior's geometric detail at the part scale and its structural coherence at the whole-object scale. As its structural input, STRUCTPARTS uses a segmented mesh scaffold, a coarse mesh paired with a face-level segmentation mask kept entirely within the prior's native 3D representation. With only lightweight finetuning, both intra-part detail and inter-part coherence emerge from the same pretrained weights. On PartObjaverse-Tiny, STRUCTPARTS matches or surpasses prior part-decomposed models at a fraction of their training cost, while naturally supporting part-mesh refinement and part-level editing.
Yiqun Zhao, Binbin Huang, Haobin Duan +2
The University of Hong Kong · Transengram · Tencent +1
Understanding and generating 3D objects as compositions of meaningful parts is fundamental to human perception and reasoning. However, most text-to-3D methods overlook the semantic and functional structure of parts. While recent part-aware approaches introduce decomposition, they remain largely geometry-focused, lacking semantic grounding and failing to model how parts align with textual descriptions or their inter-part relations. We propose DreamPartGen, a framework for semantically grounded, part-aware text-to-3D generation. DreamPartGen introduces Duplex Part Latents (DPLs) that jointly model each part's geometry and appearance, and Relational Semantic Latents (RSLs) that capture inter-part dependencies derived from language. A synchronized co-denoising process enforces mutual geometric and semantic consistency, enabling coherent, interpretable, and text-aligned 3D synthesis. Across multiple benchmarks, DreamPartGen delivers state-of-the-art performance in geometric fidelity and text-shape alignment.