Research map / Point clouds / Shape and surface completion
LPCG: A Self-conditional Architecture for Labeled Point Cloud 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 LPCG, a self-conditional labeled point cloud generator that fuses PointMAE geometry features and CLIP view features as label substitutes, trains feature diffusion modules, and uses DiT-3D as the generator, with CLIP labeling generated shapes.
Problem
Point cloud generation usually needs paired labels for conditional generation, while unconditional methods cannot annotate the generated shapes; the paper targets labeled point cloud generation without label input.
Approach
Replace label embeddings with a fused feature of self-supervised 3D geometry features (PointMAE) and CLIP view features, learn feature diffusion models to sample such conditions at inference, and use CLIP zero-shot matching over a label pool to annotate generated point clouds.
Key results
- On ShapeNet single-class generation, LPCG (DiT-3D-S, DataS) beats most baselines, outperforming LION on 10 of 12 metrics, and DiT-3D-S with DiffS on 10 of 12 metrics.
- Multi-category (chair and car) labeling accuracy with Diffusion Sampling reaches 97.44% overall (100% chair, 92.3% car).
- Using DiT-3D-L as generator yields better generation quality than DiT-3D-S in the ablation (1-NNA CD 49.60 vs 50.91).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use both geometric and view features instead of removing view features (ablation) | LPCG · ShapeNet Car, Diffusion Sampling Strategy | 1-NNA (CD): 53.48 -> 50.52 | Helped |
| Use both geometric and view features instead of geometric features alone (ablation) | LPCG · ShapeNet Car, Diffusion Sampling Strategy | COV (CD): 50.06 -> 54.16 | Helped |
| Use both geometric and view features instead of view features alone (ablation) | LPCG · ShapeNet Car, Diffusion Sampling Strategy | 1-NNA (EMD): 53.12 -> 48.56 | Helped |
| Add geometric and view features rather than use no conditional features (ablation) | LPCG · ShapeNet Car, Diffusion Sampling Strategy | COV (EMD): 50.26 -> 56.55 | Helped |
| Replace DiT-3D-S with DiT-3D-L (ablation) | LPCG · ShapeNet Car, Diffusion Sampling Strategy | COV (CD): 48.95 -> 53.12 | Helped |
| Increase feature diffusion training epochs (ablation) | LPCG · ShapeNet Car | 1-NNA (CD): 56.77 -> 48.56 | Helped |
| Increase feature diffusion training epochs (ablation) | LPCG · ShapeNet Car | 1-NNA (EMD): 54.55 -> 47.26 | Mixed |
| Increase feature diffusion training epochs (ablation) | LPCG · ShapeNet Car | COV (CD): 44.01 -> 49.73 | Helped |
| Increase feature diffusion training epochs (ablation) | LPCG · ShapeNet Car | COV (EMD): 51.56 -> 53.12 | Helped |
Surprising findings
- LPCG with DiT-3D-S reportedly beats the much larger DiT-3D-XL on 7 of 12 metrics, adding only about 6.1MB of parameters.
Limitations
- Label accuracy drops for the car class (92.3%) because CLIP misidentifies some chair features as car.
- Results on ShapeNet only; generation quality is not uniformly better than LION or DiT-3D-XL across all metrics.
Experimental setup
Datasets: ShapeNet
Benchmarks: ShapeNet
Metrics: 1-NNA COV Chamfer Distance Earth Mover's Distance label accuracy
Baselines: r-GAN PointFlow SoftFlow SetVAE DPF-Net DPM PVD LION DiT-3D-S DiT-3D-XL
Builds on: RCG DiT-3D PointMAE CLIP PointFlow
Closest papers by meaning
- Learning Progressive Point Embeddings for 3D Point Cloud Generation (CVPR 2021 · similarity 0.88)
- Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · similarity 0.87)
- Mining and Transferring Feature-Geometry Coherence for Unsupervised Point Cloud Registration (NeurIPS 2024 · similarity 0.87)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.86)
- Point-Cache: Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis (CVPR 2025 · similarity 0.86)
- Multi-View Representation is What You Need for Point-Cloud Pre-Training (ICLR 2024 · similarity 0.86)
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.