Physics-Aware Diffusion for LiDAR Point Cloud Densification

Zeping Zhang; Robert Laganière

cs.CVMar 23, 2026

Physics-Aware Diffusion for LiDAR Point Cloud Densification

Authors: Zeping Zhang, Robert Laganière

Organizations: VIVA Research Lab, Department of Electrical and Computer Engineering, University of Ottawa, Ottawa, Ontario, Canada

Abstract

LiDAR perception is severely limited by the distance-dependent sparsity of distant objects. While diffusion models can recover dense geometry, they suffer from prohibitive latency and physical hallucinations manifesting as ghost points. We propose Scanline-Consistent Range-Aware Diffusion, a framework that treats densification as probabilistic refinement rather than generation. By leveraging Partial Diffusion (SDEdit) on a coarse prior, we achieve high-fidelity results in just 156ms. Our novel Ray-Consistency loss and Negative Ray Augmentation enforce sensor physics to suppress artifacts. Our method achieves state-of-the-art results on KITTI-360 and nuScenes, directly boosting off-the-shelf 3D detectors without retraining. Code will be made available.

Figures & tables

Fig. 1: Bleeding Artifacts in LiDAR Densification. An example of a densified result generated by the baseline method LiDPM. Note that without physical constraints, existing methods often generate severe ghost artifacts extending behind objects into free space.

Fig. 2: The proposed Scanline-Consistent Range-Aware Diffusion framework. (Left) Stage-0 constructs a coarse structural prior. (Middle) Stage-1 performs Partial Diffusion. The Probabilistic Range Head predicts both range noise and an uncertainty scale $\hat{b}$ . (Right) Physics-Aware Constraints. We employ an Isotropic Probabilistic Consistency loss. The penalty is weighted by a spatial Gaussian kernel (Lateral Weight) scaled by $\hat{b}$ , allowing the uncertainty to dynamically determine the valid geometric tolerance for both lateral misalignment ( $d_{\perp}$ ) and radial occlusion ( $d_{\parallel}$ ).

Fig. 3: Negative Ray Augmentation Strategy. (Panel A) On positive rays, the model learns a probabilistic range distribution. The uncertainty $\hat{b}$ (visualized as the green confidence region) represents the isotropic spatial tolerance learned by the network. (Panel B) On negative rays sampled from known free space, we enforce a zero-occupancy constraint. This explicitly teaches the Occupancy Head to suppress ghost points in free space where the sensor beam should pass through unimpeded.

Fig. 4: Qualitative Comparison of Densification Results. LiDPM preserves local detail but introduces scattered outlier points near object edges and occasional ghost points in free space. Our method generates sharp structures with clean free space in the demonstration.

		KITTI-360 (64-beam, ${\sim}$ 120K $\to$ ${\sim}$ 250K)				nuScenes (32-beam, ${\sim}$ 34.5K $\to$ ${\sim}$ 150K)
Method	Venue	CD (m) $\downarrow$	EMD $\downarrow$	FSVR (%) $\downarrow$	REAP (%) $\downarrow$	CD (m) $\downarrow$	EMD $\downarrow$	FSVR (%) $\downarrow$	REAP (%) $\downarrow$
Non-Generative
IP-Basic [ 4 ]	IROS’18	0.2614	0.3427	4.87	10.42	0.2973	0.3846	5.21	11.63
Object-level Diffusion
PVD [ 18 ]	ICCV’21	0.3782	0.4965	11.53	17.28	0.4136	0.5374	12.74	18.95
PCDM [ 17 ]	NPL’25	0.2735	0.3581	7.64	12.35	0.3024	0.3918	8.26	13.41

TABLE I: Quantitative Comparison on KITTI-360 and nuScenes Validation Sets.

\downarrow

indicates lower is better. Best results are in bold and second-best are underlined . PVD and PCDM are capped at 8K points due to attention memory limits.

Dataset	Sparse Input	Generated (Ours)	Dense GT	Ratio
KITTI-360 (64-beam)	${\sim}$ 120K	${\sim}$ 240K	${\sim}$ 250K	2.1 $\times$
nuScenes (32-beam)	${\sim}$ 34.5K	${\sim}$ 142K	${\sim}$ 150K	4.3 $\times$

TABLE II: Dataset Statistics. Avg. point counts for sparse input (single sweep), dense ground truth (multi-sweep accumulation), and our model’s typical output.

Category	Parameter	Value
Architecture	Hidden size $H$	256
	Layers	6
	BEV Resolution	$64\times 64$
Training	Total Epochs	200
	Batch Size	8
	Optimizer	AdamW ( $lr=10^{-4}$ )

TABLE III: Implementation details of the proposed method.

Method	Ours	LiDiff ‡	LiDPM	R2DM	LiDARGen	PVD †	LiDM	IP-Basic
#Params	8.6M	38.2M	50.7M	45.5M	80.8M	13.6M	75.6M	CPU
Infer. (ms) $\downarrow$	156	2,340	1,559	434	823	OOM	20	13

TABLE IV: Computational Efficiency. We evaluate all methods with batch size 1. † PVD encountered OOM when processing 150K points.

	KITTI-360		nuScenes
Configuration	CD $\downarrow$	FSVR $\downarrow$	CD $\downarrow$	FSVR $\downarrow$	Infer. (ms)
Full Model (Ours)	0.1512	2.77	0.1671	3.96	156
w/o $\mathcal{L}_{\text{free}}$ (Ray Loss)	0.179	9.0	0.196	12.4	156
w/o Scanline Encoding	0.169	4.2	0.188	5.9	149
w/o Stage-0 (Pure Noise Init)	0.211	4.9	0.231	7.0	314
w/o BEV Branch (Point-only)	0.193	4.4	0.209	6.1	122

TABLE V: Ablation Study on KITTI-360 and nuScenes. Impact of the Ray Consistency Loss, Scanline Encoding, Coarse Initialization, and BEV Branch.

Detector	Input Data	mAP $\uparrow$	NDS $\uparrow$
VoxelNeXt	Raw Sparse (1 sweep)	61.2	68.4
	Densified (LiDPM)	64.1	70.8
	Densified (Ours)	65.6	71.9
CenterPoint	Raw Sparse (1 sweep)	56.3	64.5
	Densified (LiDPM)	59.5	66.2
	Densified (Ours)	61.1	67.6

TABLE VI: Downstream 3D Detection on nuScenes. Off-the-shelf SOTA detectors on raw sparse (1 sweep), LiDPM-densified, and our densified scans. All values are overall mAP and NDS.

3D Diffusion Models

3D diffusion models generate and reconstruct shapes, scenes, volumes, and molecular structures by gradually refining noisy 3D representations.

See all 119 papers

PointDiffusion: Diffusion-Based Scene Completion in the Point Cloud Domain
Jun 14, 2026Chidera Agbasiere, Mikhail Sannikov, Faith Ogunwoye +5Point Cloud Completion 3D Reconstruction
Intelligent Space Robotics Laboratory, Skolkovo Institute of Science and Technology, Moscow, Russia
L3DR: 3D-aware LiDAR Diffusion and Rectification
Feb 22, 2026Quan Liu, Xiaoqin Zhang, Ling Shao +13D Diffusion Models
Nanyang Technological University · Zhejiang University of Technology · UCAS-Terminus AI Lab, University of Chinese Academy of Sciences
LiDAR Resolution Recovery via Foundation-Model-Guided Diffusion
Oct 6, 2026Samed Doğan, Nico Leuze, Alfred SchöttlDiffusion Model Guidance Depth Completion
Dept. of Electrical Engineering and Information Technology Munich University of Applied Sciences, 80335 Munich, Germany

Jun 14, 2026J/K move · Enter open · S save

PointDiffusion: Diffusion-Based Scene Completion in the Point Cloud Domain

Chidera Agbasiere, Mikhail Sannikov, Faith Ogunwoye, Erik Shaikhiev+4

Intelligent Space Robotics Laboratory, Skolkovo Institute of Science and Technology, Moscow, Russia

Point Cloud Completion 3D Reconstruction 3D Scene Completion Diffusion Model Inference Acceleration Diffusion Models

Open paper AskPDF arXiv

Reconstructing dense 3D scenes from sparse LiDAR point clouds is a fundamental challenge in autonomous driving, where latent diffusion models offer a promising solution. However, existing approaches rely on object-level autoencoders that collapse into unstable global representations at outdoor scale and suffer from ground truth data corrupted by odometry drift that systematically degrades supervision quality. Furthermore, multi-step diffusion inference incurs prohibitive latency for real-time deployment. We propose a novel multi-token Gaussian VAE with cross-attention pooling for stable scene-scale LiDAR compression, combined with an anchor-based ICP ground truth refinement pipeline that eliminates drift-induced noise from training supervision. Together, these components enable a scaffold-free single-step diffusion completion model that achieves an approximately 16x reduction in squared Chamfer distance on SemanticKITTI seq. 08 (0.396 m^2 to 0.024 m^2), surpasses LiDiff and ScoreLiDAR by 17-19% and 10-11%, respectively, and operates at 25-143x lower inference latency. Our results demonstrate that data quality dominates model design in this regime and that multi-token latent spaces provide a stable first stage for latent diffusion-based scene completion.

Abstract

3D Diffusion Models

Explore similar work

PointDiffusion: Diffusion-Based Scene Completion in the Point Cloud Domain

L3DR: 3D-aware LiDAR Diffusion and Rectification

LiDAR Resolution Recovery via Foundation-Model-Guided Diffusion