New diffusion-based method speeds up LiDAR point cloud densification to 156ms while eliminating ghost points through physics-aware constraints

Scanline-Consistent Range-Aware Diffusion framework treats point cloud densification as probabilistic refinement rather than generation, achieving high-fidelity results in just 156 milliseconds

Novel Ray-Consistency loss and Negative Ray Augmentation techniques enforce sensor physics to suppress physical hallucinations and artifact ghost points that plague existing diffusion models

Achieves state-of-the-art performance on KITTI-360 and nuScenes benchmarks and directly improves off-the-shelf 3D detectors without requiring model retraining

Addresses critical LiDAR limitation where distant objects suffer from distance-dependent sparsity that severely impacts perception accuracy