Two Steps Are All You Need: Efficient 3D Point Cloud Anomaly Detection with Consistency Models

A team of researchers (Pranav A, Shashank B, Pranav Siddappa, Dominik Seuss, Minal Moharir, Subramanya KN) has developed a new approach to 3D point cloud anomaly detection that dramatically reduces inference cost. Traditional diffusion models require hundreds of iterative denoising steps to reconstruct anomaly-free geometry, making them impractical for resource-constrained, latency-critical systems. The proposed method reformulates reconstruction-based anomaly detection through consistency learning, enabling direct prediction of clean 3D data in one or two network evaluations.

In benchmarks, the new model achieves up to 80x faster runtime than the current state-of-the-art method (R3D-AD) without any GPU acceleration, while preserving strong detection performance: 76.20% I-AUROC on Anomaly-ShapeNet and 72.80% I-AUROC on Real3D-AD. The approach introduces a novel hybrid loss that explicitly enforces reconstruction toward clean data, further boosting reliability. This breakthrough makes high-speed 3D anomaly detection feasible for edge devices like drones and smart industrial cameras, where every millisecond counts.