L3DR: 3D-aware LiDAR Diffusion and Rectification

Quan Liu; Xiaoqin Zhang; Ling Shao; Shijian Lu

cs.CVFeb 22, 2026

L3DR: 3D-aware LiDAR Diffusion and Rectification

Authors: Quan Liu, Xiaoqin Zhang, Ling Shao, Shijian Lu

Organizations: Nanyang Technological University · Zhejiang University of Technology · UCAS-Terminus AI Lab, University of Chinese Academy of Sciences

Abstract

Range-view (RV) based LiDAR diffusion has recently made huge strides towards 2D photo-realism. However, it neglects 3D geometry realism and often generates various RV artifacts such as depth bleeding and wavy surfaces. We design L3DR, a 3D-aware LiDAR Diffusion and Rectification framework that can regress and cancel RV artifacts in 3D space and restore local geometry accurately. Our theoretical and empirical analysis reveals that 3D models are inherently superior to 2D models in generating sharp and authentic boundaries. Leveraging such analysis, we design a 3D residual regression network that rectifies RV artifacts and achieves superb geometry realism by predicting point-level offsets in 3D space. On top of that, we design a Welsch Loss that helps focus on local geometry and ignore anomalous regions effectively. Extensive experiments over multiple benchmarks including KITTI, KITTI360, nuScenes and Waymo show that the proposed L3DR achieves state-of-the-art generation and superior geometry-realism consistently. In addition, L3DR is generally applicable to different LiDAR diffusion models with little computational overhead.

Figures & tables

Figure 2: Empirical validation of Theorem 1 . The graph shows the distribution of

\|\nabla x\|

for GT, vanilla RV diffusion, and our rectified RV, including the corresponding Jensen-Shannon Divergence (JSD) w.r.t. the GT.

Figure 3: The training pipeline of the proposed L3DR framework. In the LiDAR diffusion training stage, generated and ground-truth point cloud pairs are collected using semantic-conditioned LiDAR diffusion. In the residual regression training stage, such data pairs are employed to train a 3D network to remove RV artifacts present in the residuals to improve generation quality.

Figure 4: Visualization of two types of errors in RRN training data. While the generated point clouds ( c o l or ed ) approximate the GT ( gray ) in most of the regions with high-variance errors , i.e., RV artifacts as highlighted with green dotted lines, there are also regions with high-bias errors which impede training, including (1) shifted walls, (2) random points on the leaves where laser hits are hard to predict, and (3) isolated chunks with consistent depth error. These bias-dominated regions are harmful for RRN training.

Figure 5: The inference pipeline of the proposed L3DR.

		Perceptual		Distributional
Task	Method	FSVD $\downarrow$	FPVD $\downarrow$	JSD $\downarrow$	MMD $\times 10^{(-4)}\downarrow$
KITTI360 Unconditional	LiDARGAN [ 3 ]	183.4	168.1	0.272	4.74
	LiDARVAE [ 3 ]	129.9	105.8	0.237	7.07
	ProjectedGAN [ 36 ]	44.7	33.4	0.188	2.88
	UltraLiDAR [ 47 ]	72.1	66.6	0.747	17.12
	LiDARGen(1160s) [ 57 ]	39.2	33.4	0.188	2.88

Table 1: Benchmarking of unconditional generation on KITTI360 and semantic-conditioned generation on nuScenes and Waymo. For the semantic-conditioned experiments, RRN takes segmentation map as additional input for optimal performance. Gray areas highlight direct comparisons with the baselines.

	Semantic-Map-Conditioned				Front-Image-Conditioned
Method	FSVD $\downarrow$	FPVD $\downarrow$	JSD $\downarrow$	MMD $\times 10^{(-4)}\downarrow$	FSVD $\downarrow$	FPVD $\downarrow$	JSD $\downarrow$	MMD $\times 10^{(-4)}\downarrow$
LiDARGen	31.7	30.1	0.130	5.18	-	-	-	-
LDM	21.3	20.3	0.088	3.73	35.9	26.5	0.26	3.80
LiDM	20.2	17.7	0.072	3.16	32.5	25.8	0.21	3.69
Ours	20.4	15.0	0.070	1.69	33.6	24.9	0.18	3.30
$\Delta$ Inc.	-1.0%	+15.3%	+2.8%	+46.5%	-3.4%	+3.5%	+12.2%	+10.6%

Table 2: Comparison of conditional LiDAR point cloud generation on SemanticKITTI and KITTI360. Gray areas highlight direct comparisons with the baseline, LiDM. ‘Ours-Sem’ denotes our method with segmentation input to the RRN.

Figure 6: Visualization of conditional generation on SemanticKITTI. Cyan regions highlight the improved RV artifacts from the diffusion-generated (i.e., denoised) data to our rectified data, such as depth bleeding and wavy surfaces.

Configurations			Perceptual		Statistical
Backbone	Loss	RRN Sem. Input	FSVD $\downarrow$	FPVD $\downarrow$	JSD $\times 10^{(-2)}\downarrow$	MMD $\times 10^{(-5)}\downarrow$
/ (baseline)	/	-	18.3	15.3	7.1	16.2
SPUNet	Welsch	-	16.4	12.1	6.7	16.7
SPUNet	MSE	-	26.3	25.1	7.0	12.6
SPUNet	Welsch	✓	12.5	10.7	6.7	15.0
PTV3	Welsch	-	17.7	13.9	6.8	17.2

Table 3: Ablation experiment on SemanticKITTI, including RRN backbone structure, loss function, semantic-map input to RRN, and a fair baseline using a 2D image Unet instead of a 3D UNet.

Method	Network	# Params (M)	Steps	Time (ms)
R2DM	Eff-UNet	31.10	50	579.83
LiDM	VQ-VAE, UNet	257.77	50	557.36
+Ours	SPUNET	+37.90	+1	+19.65

Table 4: Computational overhead on KITTI360. our method introduce very slight computational overhead over the baselines.

Figure 7: Visualization of unconditional generation on KITTI360. Red regions highlight the the improvements from the diffusion-generated (i.e., denoised) data to our rectified data.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: Intuitive validation for Theorem 1 , i.e., 2D and 3D models have inherently different sensitivity to disturbances. (a) Adding large noises ( $\sigma=5$ m) does not hinder human understanding of image contents on RV, yet completely sets the point cloud into a mess. (b) Adding small noises ( $\sigma=0.2$ m) adds visible noises in the point cloud while retaining its spatial structure, but this noise scale is almost non-percievable in RV.

Configurations			Perceptual		Statistical
Backbone	Loss	RRN Sem. Input	FSVD $\downarrow$	FPVD $\downarrow$	JSD $\times 10^{(-2)}\downarrow$	MMD $\times 10^{(-5)}\downarrow$
/ (baseline)	/	-	11.7	13.1	11.6	21.5
SPUNet	Welsch	-	10.6	12.1	6.7	16.6
SPUNet	MSE	-	27.0	25.7	10.1	22.3
SPUNet	Welsch	✓	9.8	9.7	9.5	20.9
PTV3	Welsch	-	9.3	9.8	9.7	22.1

Appendix

Table 5: Ablation experiment on nuScenes, including RRN backbone structure, loss function and semantic-map input to RRN.

Configurations			Perceptual		Statistical
Backbone	Loss	RRN Sem. Input	FSVD $\downarrow$	FPVD $\downarrow$	JSD $\times 10^{(-2)}\downarrow$	MMD $\times 10^{(-5)}\downarrow$
/ (baseline)	/	-	11.8	12.4	9.5	13.4
SPUNet	Welsch	-	9.8	9.9	7.9	13.4
SPUNet	MSE	-	24.8	27.3	8.7	11.7
SPUNet	Welsch	✓	9.6	11.6	7.8	13.0
PTV3	Welsch	-	10.3	11.4	7.6	13.6

Appendix

Table 6: Ablation experiment on Waymo, including RRN backbone structure, loss function, and semantic-map input to RRN.

$\nu$	Perceptual		Statistical
$\nu$	FSVD $\downarrow$	FPVD $\downarrow$	JSD $\times 10^{(-2)}\downarrow$	MMD $\times 10^{(-5)}\downarrow$
/ (baseline)	18.3	15.3	7.1	16.2
0.01	18.1	14.9	6.9	16.2
0.05	17.9	14.3	6.9	16.3
0.1	17.9	13.4	7.0	16.3
0.5	17.7	13.9	6.8	17.2

Appendix

Table 7: Parameter tuning experiment of

\nu

with PTV3 on SemanticKITTI.

RRN Training Data	Test Data	C-D $\downarrow$	F-score $\uparrow$	RMSE $\downarrow$	MedAE $\downarrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$
Val (3 frames $\times$ 7 scenes)	Val (1 frame $\times$ 7 scenes)	0.0557	0.966	1.64	0.00647	0.0233	0.978	33.9
None (GS-LiDAR original)	Val (1 frame $\times$ 7 scenes)	0.0560	0.963	1.64	0.00832	0.0231	0.978	34.0
Train (46 frames $\times$ 7 scenes)	Val (4 frames $\times$ 7 scenes)	0.0473	0.966	2.46	0.0171	0.0651	0.776	30.3
None (GS-LiDAR original)	Val (4 frames $\times$ 7 scenes)	0.0480	0.961	2.21	0.0156	0.0518	0.818	31.2

Appendix

Table 8: RRN performance on GS-LiDAR reconstructed LiDAR data under different training and test data configurations.

Method	Benefits	Limitations
RV Diffusion (e.g., R2DM, LiDM, ours)	• Lightweight and low-cost generation	• Single-frame generation Geometry artifacts • Moderate Gen2Real gap))
	• Broad condition compatibility
	• Natural modeling of LiDAR’s projective structure
BEV Diffusion (e.g., UltraLiDAR)	• Supports manipulation of BEV grid • Local spatial control with partial point input	• Poor projection realism
		• Limited multimodal control
		• Still single-frame

Appendix

Table 9: Comparative overview of existing genres for LiDAR point cloud generation.

Figure 9: Visualization of conditional generation on nuScenes. Our results exhibit better geometry-realism such as sharp borders, flat surfaces, less noise points and reduced depth bleeding.

Figure 10: Visualization of conditional generation on Waymo Open Dataset. Our results exhibit better geometry-realism with reduced depth bleeding, flattened surfaces, rectified corners and reduced noise points.

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

Physics-Aware Diffusion for LiDAR Point Cloud Densification
Mar 23, 2026Zeping Zhang, Robert Laganière3D Diffusion Models Physics-Informed ML
VIVA Research Lab, Department of Electrical and Computer Engineering, University of Ottawa, Ottawa, Ontario, Canada
Geometry Without Coordinates: LiDAR Diffusion as a 3D Feature Bridge
Sep 9, 2026Samed Doğan, Nico Leuze, Alfred Schöttl3D Semantic Segmentation Transfer Learning
Department of Electrical Engineering and Information Technology Munich University of Applied Sciences Munich, Bavaria 80335
LiFi: LiDAR Generation from Multi-View Images via Geometric and Semantic Collaborative Guidance
Oct 14, 2026Sizhuo Zhou, Xiaosong Jia, Fanrui Zhang +10Point Cloud Generation Diffusion Model Guidance
University of Science and Technology of China · Shanghai Innovation Institute · Fudan University +1

Mar 23, 2026J/K move · Enter open · S save

Physics-Aware Diffusion for LiDAR Point Cloud Densification

Zeping Zhang, Robert Laganière

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

3D Diffusion Models Physics-Informed ML

Open paper AskPDF arXiv

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.

Abstract

3D Diffusion Models

Explore similar work

Physics-Aware Diffusion for LiDAR Point Cloud Densification

Geometry Without Coordinates: LiDAR Diffusion as a 3D Feature Bridge

LiFi: LiDAR Generation from Multi-View Images via Geometric and Semantic Collaborative Guidance