Research map / Point clouds / Shape and surface completion
SeaLion: Semantic Part-Aware Latent Point Diffusion Models for 3D Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SeaLion, a latent point diffusion model that jointly denoises latent points and part segmentation labels, plus a part-aware Chamfer distance (p-CD) for evaluating segmentation-labeled point cloud generation.
Problem
Generating 3D point clouds with point-wise part segmentation labels, and evaluating such labeled generations, since CD/EMD-based metrics ignore segmentation and existing part-wise metrics miss inter-part coherence.
Approach
Uses a hierarchical latent point diffusion model (building on Lion) whose point-level U-Net has a shared down-sampling path and two parallel up-sampling branches that jointly predict noise and segmentation labels, conditioning the decoder on the labels; and defines p-CD, a Chamfer distance over all parts jointly (infinite if part sets differ) so 1-NNA/COV/MMD measure inter-part coherence.
Key results
- SeaLion outperforms DiffFacto by an average of 13.33% on 1-NNA (p-CD) across airplane, car, chair, and lamp on ShapeNet.
- On IntrA, SeaLion improves over DiffFacto by 6.52% on 1-NNA (p-CD) (65.22% vs 71.74%) and 21.74% on COV (p-CD).
- Semi-supervised training with 10% labeled car data plus unlabeled data lowers 1-NNA (p-CD) from 90.82 (DiffFacto) to 83.23.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SeaLion instead of DiffFacto | ShapeNet airplane generation · ShapeNet airplane | 1-NNA (p-CD) (%): 81.67 -> 65.40 | Helped |
| Use SeaLion instead of DiffFacto | ShapeNet car generation · ShapeNet car | 1-NNA (p-CD) (%): 90.51 -> 73.10 | Helped |
| Use SeaLion instead of DiffFacto | ShapeNet chair generation · ShapeNet chair | 1-NNA (p-CD) (%): 77.34 -> 63.14 | Helped |
| Use SeaLion instead of DiffFacto | IntrA aneurysm generation · IntrA aneurysm | 1-NNA (p-CD) (%): 71.74 -> 65.22 | Helped |
| Add unlabeled training data to SeaLion trained with 10% labeled data (ablation) | SeaLion · ShapeNet car | 1-NNA (p-CD) (%): 87.34 -> 83.23 | Helped |
| Add SeaLion-generated samples to the traditionally augmented training set | SPoTr · ShapeNet airplane | mIoU: 82.55 -> 83.81 | Helped |
| Add SeaLion-generated samples to the traditionally augmented training set | SPoTr · ShapeNet car | mIoU: 78.09 -> 79.43 | Helped |
| Add SeaLion-generated samples to the traditionally augmented training set | SPoTr · ShapeNet chair | mIoU: 90.83 -> 90.88 | Helped |
| Add SeaLion-generated samples to the traditionally augmented training set | SPoTr · ShapeNet guitar | mIoU: 91.07 -> 91.56 | Helped |
| Add SeaLion-generated samples to the traditionally augmented training set | SPoTr · ShapeNet lamp | mIoU: 83.18 -> 84.54 | Helped |
| Add SeaLion-generated samples to the traditionally augmented training set | SPoTr · ShapeNet table | mIoU: 82.48 -> 83.44 | Helped |
Surprising findings
- DiffFacto scores well on the intra-part 1-NNA-P metric but drops sharply under p-CD, showing that per-part metrics miss implausible inter-part assemblies.
Limitations
- SeaLion is trained separately per category rather than as a single multi-category model.
- Chamfer-based p-CD is defined as infinite for shapes with different part sets, so it only applies to matching part structures.
- Evaluated on only six ShapeNet categories plus one medical dataset.
Experimental setup
Datasets: ShapeNet IntrA
Benchmarks: ShapeNet IntrA
Metrics: 1-NNA (p-CD) COV (p-CD) MMD (p-CD) 1-NNA-P COV-P MMD-P mIoU
Baselines: DiffFacto Lion & PointNet++ Lion & SPoTr
Builds on: Lion DiffFacto PVCNN DDPM
Closest papers by meaning
- PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.88)
- TIGER: Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process (CVPR 2024 · similarity 0.86)
- PartGen: Part-level 3D Generation and Reconstruction with Multi-view Diffusion Models (CVPR 2025 · similarity 0.86)
- GPLD3D: Latent Diffusion of 3D Shape Generative Models by Enforcing Geometric and Physical Priors (CVPR 2024 · similarity 0.86)
- PointInfinity: Resolution-Invariant Point Diffusion Models (CVPR 2024 · similarity 0.86)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.