Research map / Point clouds / Shape and surface completion
ComPC: Completing a 3D Point Cloud with 2D Diffusion Priors
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 a test-time, training-free point cloud completion framework that renders partial points with 3D Gaussian Splatting and uses Zero-1-to-3 diffusion priors (SDS) to infer missing geometry, then extracts uniform points.
Problem
Point clouds captured by sensors are incomplete due to self-occlusion, and supervised completion networks fail to generalize to unseen categories and real scans; prior test-time method SDS-Complete needs text prompts and is very slow.
Approach
Render a reference image from an estimated most-complete viewpoint of the partial point cloud (Gaussians with unit opacity and normal-map colors), then optimize extra 3D Gaussians with SDS from Zero-1-to-3 conditioned on that image, with a Chamfer-based preservation constraint, binarized opacity, and a grid-pulling resampling step to extract uniform points, requiring no text prompts or training.
Key results
- On synthetic point clouds, ComPC averages CD/EMD 1.43/1.88 versus 3.68/4.44 for the best-listed network baseline (ShapeFormer) in the synthetic table.
- On Redwood real scans, ComPC averages CD 1.35 (in-domain) and 2.57 (out-of-domain), against SDS-Complete's 2.15 and 3.33.
- Completing one Redwood point cloud takes about 15 minutes on an RTX A6000, versus up to 1950 minutes for SDS-Complete.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use normal-map colorization instead of depth colorization (ablation) | ComPC Partial Gaussian Initialization · Redwood in-domain and out-domain average | CD: 2.25 -> 1.96 | Helped |
| Use normal-map colorization instead of coordinate colorization (ablation) | ComPC Partial Gaussian Initialization · Redwood in-domain and out-domain average | CD: 2.01 -> 1.96 | Helped |
| Add Preservation Constraint to view-dependent guidance (ablation) | ComPC Zero-shot Fractal Completion · Redwood | EMD: 3.20 -> 3.07 | Helped |
| Add Preservation Constraint to view-dependent guidance (ablation) | ComPC Zero-shot Fractal Completion · Synthetic | CD: 3.35 -> 2.55 | Helped |
| Add Gaussian Surface Extraction (ablation) | ComPC Point Cloud Extraction · Redwood | CD: 1.97 -> 1.50 | Helped |
| Add Gaussian Surface Extraction (ablation) | ComPC Point Cloud Extraction · Redwood | EMD: 3.07 -> 3.38 | Hurt |
| Add Grid Pulling after surface extraction (ablation) | ComPC Point Cloud Extraction · Redwood | EMD: 3.38 -> 2.60 | Helped |
| Add Grid Pulling after surface extraction (ablation) | ComPC Point Cloud Extraction · Redwood | CD: 1.50 -> 1.96 | Hurt |
| Add merge layer to Grid Pulling (ablation) | ComPC Point Cloud Extraction · Redwood | CD: 2.04 -> 1.96 | Helped |
| Add the near-point loss to Grid Pulling (ablation) | ComPC Point Cloud Extraction · Redwood | CD: 3.11 -> 2.04 | Helped |
| Use shared scalar scaling (ablation) | ComPC 3D Gaussian modifications · Redwood | CD: 2.50 -> 1.96 | Helped |
| Use binary opacity (ablation) | ComPC 3D Gaussian modifications · Redwood | CD: 6.35 -> 2.50 | Helped |
Surprising findings
- The method completes a plausible teapot handle without any text prompt, suggesting the diffusion prior recognizes object category from the reference image.
Limitations
- Test-time optimization still takes around 15 minutes per object.
- Thin surfaces in occluded regions are hallucinated as thicker structures by the 2D diffusion prior.
- Performance degrades with input noise (e.g., CD rises to 3.18 at noise std 0.01 on synthetic data).
Experimental setup
Datasets: Redwood synthetic objects from established 3D model sources
Benchmarks: Redwood synthetic point clouds ShapeNet
Metrics: Chamfer Distance Earth Mover's Distance TMD UHD MMD
Baselines: PoinTr SeedFormer PointAttN ShapeFormer SVDFormer AdaPoinTr SDS-Complete
Builds on: 3D Gaussian Splatting Zero-1-to-3 SDS-Complete DreamGaussian NeuralPull
Closest papers by meaning
- Digging into Intrinsic Contextual Information for High-fidelity 3D Point Cloud Completion (AAAI 2025 · similarity 0.90)
- UniPre3D: Unified Pre-training of 3D Point Cloud Models with Cross-Modal Gaussian Splatting (CVPR 2025 · similarity 0.90)
- Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians (CVPR 2025 · similarity 0.89)
- SuperPC: A Single Diffusion Model for Point Cloud Completion, Upsampling, Denoising, and Colorization (CVPR 2025 · similarity 0.89)
- GaussianPainter: Painting Point Cloud into 3D Gaussians with Normal Guidance (AAAI 2025 · similarity 0.88)
- PCDreamer: Point Cloud Completion Through Multi-view Diffusion Priors (CVPR 2025 · similarity 0.88)
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.