Research map / Point clouds / Shape and surface completion

GenPC: Zero-shot Point Cloud Completion via 3D Generative Priors

CVPR 2025 · Point clouds › Shape and surface completion · approach: Point cloud completion networks · 7 citations (OpenAlex)

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.

Zero-shot point cloud completion by prompting a feed-forward image-to-3D generative model with depth images rendered from partial scans, then fusing the generated shape with the input via scale/pose alignment and optional SDS refinement.

Problem

Point cloud completion methods trained on synthetic datasets generalize poorly to out-of-distribution real-world scans and unseen categories, and prior zero-shot SDS-based methods are slow and yield coarse details.

Approach

Uses explicit 3D priors from a feed-forward image-to-3D generative model instead of per-object SDS optimization: a Depth Prompting module converts a partial point cloud into an image via viewpoint selection (hidden point removal), depth inpainting, and ControlNet; a Geometric Preserving Fusion module then aligns the generated shape to the input by searching scale with ICP using XYZ and RGB Chamfer

Key results

• On Redwood, GenPC reaches average CD 1.74 and EMD 2.88 (x10^2), versus SDS-Complete 2.72/4.06, a 36% CD reduction and 29% EMD reduction.
• On ScanNet (tables and chairs), GenPC averages CD 1.62 and EMD 3.55, beating SnowflakeNet (1.74/4.27) and AdaPoinTr (2.21/2.38 in CD/EMD columns as reported).
• Ablations show removing the 3D generative model (Gaussian noise init) raises Redwood CD from 1.74 to 4.65, and removing dynamic scale adaptation raises it to 4.38.

Measured effects reported in the paper

Surprising findings

• Without depth inpainting, the method performs well on dense Redwood scans but drops sharply on sparse ScanNet inputs.
• The optional SDS refinement stage gives only modest gains; the merged point cloud before refinement often performs competitively, so refinement was made optional for speed.

Limitations

• Refinement via SDS is time-consuming and optional; the base pipeline depends on the quality of depth projection and inpainting for sparse inputs.
• Evaluation is limited to a few categories (tables, chairs on ScanNet; nine Redwood object classes) and qualitative KITTI results only.

Experimental setup

Datasets: Redwood ScanNet KITTI

Benchmarks: Redwood ScanNet KITTI

Metrics: Chamfer Distance Earth Mover's Distance

Baselines: PoinTr SnowflakeNet AdaPoinTr SDS-Complete

Builds on: SDS-Complete Zero123 ControlNet LGM InstantMesh DreamFusion

Closest papers by meaning

• ComPC: Completing a 3D Point Cloud with 2D Diffusion Priors (ICLR 2025 · similarity 0.87)
• Point-based Instance Completion with Scene Constraints (ICLR 2025 · similarity 0.87)
• PCDreamer: Point Cloud Completion Through Multi-view Diffusion Priors (CVPR 2025 · similarity 0.87)
• Position-Aware Guided Point Cloud Completion with CLIP Model (AAAI 2025 · similarity 0.87)
• SuperPC: A Single Diffusion Model for Point Cloud Completion, Upsampling, Denoising, and Colorization (CVPR 2025 · similarity 0.86)
• Point2Real: Bridging the Gap between Point Cloud and Realistic Image for Open-World 3D Recognition (AAAI 2024 · similarity 0.86)