HuLiGen: Human LiDAR Generation from Parametric Body Models

Salma Galaaoui; Nermin Samet; David Picard

cs.CVOct 7, 2026

HuLiGen: Human LiDAR Generation from Parametric Body Models

Authors: Salma Galaaoui, Nermin Samet, David Picard

Organizations: Valeo.ai, Paris, France · LIGM, CNRS, Univ Gustave Eiffel, ENPC, Institut Polytechnique de Paris, Marne-la-Vall´ee, France

Abstract

LiDAR point clouds of humans are extremely expensive to collect and annotate, thus represent a scarce resource that hinders the development of human analysis using this modality. To alleviate this scarcity, prior work relies on simulated human LiDAR, but such samples do not fully reflect the geometry and sensing characteristics of real observations. In contrast, we introduce HuLiGen, a generative model that generates human LiDAR point clouds from a parametric body model, using a point transformer trained with a flow-matching objective. We show that our generated point clouds are closer to the real capture distribution. Using HuLiGen to generate synthetic data, we propose a synthetic-only pretraining scheme for LiDAR-based HPE that achieves state-of-the-art performance, with even larger gains in low-annotation and low-data regimes, where MPJPE is reduced by up to 50%. Code, models and generated samples are available at https://github.com/valeoai/HuLiGen.

Figures & tables

Panel of Figure 2 — Figure 2 : Root-joint axes across different pose types from the SLOPER4D [ 7 ] dataset. We use the azimuth of the lateral axis for canonicalisation, as its projection onto the horizontal plane remains well defined across the different poses. In contrast, the forward axis can approach the vertical under strong hip flexion, causing its horizontal projection to shrink (at the risk of reaching 0 ∘ ) and making the corresponding azimuth ill-conditioned.

Figure 3 : Our synthetic pretraining pipeline. We sample local body poses

\theta

from AMASS and bootstrap the remaining conditioning parameters,

\mathbf{o}

\beta

as well as number of target points from the target dataset. These conditions are provided to our flow-matching model to generate novel human LiDAR observations. We then pretrain the HPE model on the generated samples and finetune it on real target dataset, yielding substantial performance gains.

Canonical align.	Shape param.	Cartesian joints	CD $\downarrow$	1-NNA $\downarrow$	COV $\uparrow$	FPD $\downarrow$	KPD $\downarrow$
✗	✗	✗	0.02983	90.65	38.93	0.1857	0.01357
✓	✗	✗	0.02836	90.46	40.21	0.1890	0.01689
✗	✓	✗	0.02859	91.43	39.20	0.1725	0.01560
✓	✓	✗	0.02718	88.11	43.99	\B 0.1277	0.01273
✓	✓	✓	0.02648	\B 84.45	\B 45.17	0.1508	\B 0.01124
	Raycasting		\B 0.02470	95.45	37.77	0.1390	0.01177

Table 1 : Conditioning ablation study on SLOPER4D using generative quality metrics . Canonical align.: ✓ = canonical axis transformation applied; ✗ = no canonical transformation. Shape param.: ✓ = shape parameters included in the conditioning; ✗ = shape parameters excluded from the conditioning. Cartesian joints: ✓ = Cartesian joint coordinates with PointNet embeddings; ✗ = 6D rotation coordinates with cyclic embeddings.

Method	CD $\downarrow$	1-NNA $\downarrow$	COV $\uparrow$	FPD $\downarrow$	KPD $\downarrow$
Raycasting	0.0513	94.23	22.49	0.6051	0.07869
HuLiGen	0.0566	75.64	37.55	0.0961	0.00422

Table 2 : Generative quality comparison on WOD.

Dataset	Syn. pre. data size	Pretraining		Finetuning		$\mathbf{-\Delta\%}$
Dataset	Syn. pre. data size	MPJPE	PA-	MPJPE	PA-	$\mathbf{-\Delta\%}$
SLOPER4D	N/A	N/A	N/A	6.02 ∗	4.58 ∗	N/A
	8k	15.13	9.31	5.18	4.07	13.95
	24k	12.60	7.63	5.05	4.02	16.11
	48k	\B 11.09	\B 7.03	\B 5.04	\B 3.97	16.28
WOD	N/A	N/A	N/A	6.74 ∗	5.09 ∗	N/A

Table 3 : Effect of synthetic pretraining on LiDAR-based 3D HPE. ∗ denotes results obtained by training on real data from scratch, without pretraining.

\Delta

values report the change in MPJPE relative to training from scratch without pretraining.

Dataset	Method	Pretraining		Finetuning		$\mathbf{-\Delta\%}$
Dataset	Method	MPJPE	PA-	MPJPE	PA-	$\mathbf{-\Delta\%}$
SLOPER4D	HuLiGen	11.09	\B 7.03	\B 5.04	\B 3.97	\B 16.28
	Rc — bootstr.	\B 11.03	7.26	5.47	4.26	9.14
	Rc — direct	21.95	9.73	5.27	4.16	12.45
WOD	HuLiGen	\B 10.79	\B 8.51	\B 5.85	\B 4.37	\B 13.20
	Rc — bootstr.	15.54	10.39	5.98	4.42	11.28

Table 4 : Comparison of raycasted and HuLiGen -generated samples using our pretraining pipeline on the downstream LiDAR-based 3D HPE task.

\Delta

values report the change in MPJPE relative to training from scratch without pretraining.

Table 5 : SOTA comparison for LiDAR-based 3D HPE. Bench.: ✓ = the benchmark data is the only real-data source used in the pipeline; ✗ = training includes extra large-scale real data. LiDAR: ✓ = input is only LiDAR point clouds; ✗ = input includes images.

Method	Pt.	Zero-shot (0%)		1 %		10 %
Method	Pt.	MPJPE	PA-	MPJPE	PA-	MPJPE	PA-
Random init.	✗	N/A	N/A	18.01	14.19	13.74	10.60
Rc — bootstr.	✓	15.54	10.39	9.01	7.15	7.22	5.45
Rc — direct	✓	15.11	10.04	10.89	8.37	7.98	6.16
GC-KPL [ 40 ]	✓	17.70	-	18.52	-	15.10	-
HuLiGen	✓	\B 10.79	\B 8.51	\B 8.43	\B 6.48	\B 6.84	\B 5.18

Table 6 : Results in the low-annotation regime on the WOD validation set. Zero-shot denotes evaluation directly using models pretrained on synthetic data, without finetuning on real data. Pt. denotes pretraining.

Fraction	Pt.	Ft.	SLOPER4D		$\mathbf{-\Delta\%}$	WOD		$\mathbf{-\Delta\%}$
Fraction	Pt.	Ft.	MPJPE	PA-		MPJPE	PA-
10%	✗	✓	$9.292$	$6.618$		$13.273$	$10.267$
	✓	✗	$16.398$	$9.376$		$14.072$	$8.857$
	✓	✓	\B 6.464	\B 4.863	\B 30.46	\B 7.568	\B 5.614	\B 42.95
20%	✗	✓	$7.906$	$5.603$		$10.526$	$8.281$
	✓	✗	$13.825$	$8.646$		$13.775$	$8.773$

Table 7 : Results in the low-data regime on the WOD and SLOPER4D validation sets using 10% and 20% of the real training data. Pt: pretraining; Ft: finetuning.

Panel of Figure 4 — Figure 4 : Qualitative results on the WOD [ 39 ] and SLOPER4D [ 7 ] validation sets. For each sample, we show the camera crop, body-model surface, real LiDAR observation, raycasted result and our generation, with point counts below. Camera images are for visualization only, and all panels share the same viewpoint and scale. Ray casting stays on the bare body surface, while real and generated observations capture off-surface structure from clothing and hair.

Panel of Figure 5 — Figure 5 : Qualitative results on AMASS. AMASS provides pose and shape but no sensor data, so we use the body surface as reference. For each sample, we show the body surface, raycast result and our generation, with point counts below. Ray casting remains on the bare body surface, yielding a median surface distance of 0 cm on both datasets. In contrast, our generations reach 1.41 cm on SLOPER4D (top row), close to 1.17 cm for real observations, and 4.5 cm on WOD compared with 3.0 cm for real data. We attribute the larger WOD (bottom row) gap mainly to ground-level noise.

Canonical align.	Shape param.	Cartesian joints	$\mathbf{o}_{rep}$	$\theta_{rep}$	$o_{emb}$	$\beta_{emb}$	$\theta_{emb}$	token count
✗	✗	✗	abs. cyl.	24 $\times$ 6D rotation	4	N/A	144	28
✓	✗	✗	canon.	24 $\times$ 6D rotation	3	N/A	144	27
✗	✓	✗	abs. cyl.	24 $\times$ 6D rotation	4	10	144	38
✓	✓	✗	canon.	24 $\times$ 6D rotation	3	10	144	37
✓	✓	✓	canon.	24 $\times$ 3D position	3	10	72	37

Table 8 : Conditioning and embedding details . Canonical align.: ✓ = canonical axis transformation applied; ✗ = no canonical transformation. Shape param.: ✓ = shape parameters included in the conditioning; ✗ = shape parameters excluded from the conditioning. Cartesian joints: ✓ = Cartesian joint coordinates with PointNet embeddings; ✗ = 6D rotation coordinates with cyclic embeddings. abs. cyl. denotes absolute cylindrical coordinates and canon. denotes canonicalized coordinates. The third column group reports the dimensionality of each conditioning representation before embedding; while the final column gives the total conditioning-token count used by each variant.

Explore similar work

LaGen: Towards Autoregressive LiDAR Scene Generation
Nov 26, 2025Sizhuo Zhou, Xiaosong Jia, Fanrui Zhang +73D Scene Generation Autoregressive Generation
University of Science and Technology of China, Hefei, Anhui 230026, China · Shanghai Innovation Institute, Shanghai 200231, China · Fudan University, Shanghai 200433, China +1
OmniLiDAR: A Unified Diffusion Framework for Multi-Domain 3D LiDAR Generation
May 13, 2026Youquan Liu, Weidong Yang, Ao Liang +9Synthetic Data Generation for Robotics 3D Scene Generation
College of Computer Science and Artificial Intelligence, Fudan University, Shanghai, China · School of Computing, Department of Computer Science, National University of Singapore, Singapore · College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing, China +5
TAP3D: Thermal-Assisted 3D Human Point Clouds
Oct 8, 2026Xie Zhang, Chengxiao Li, Xuan Liu +1

Nov 26, 2025, cs.CVJ/K move · Enter open · S save

LaGen: Towards Autoregressive LiDAR Scene Generation

Sizhuo Zhou, Xiaosong Jia, Fanrui Zhang, Junjie Li+6

University of Science and Technology of China, Hefei, Anhui 230026, China · Shanghai Innovation Institute, Shanghai 200231, China · Fudan University, Shanghai 200433, China+1

3D Scene Generation Autoregressive Generation World Models for Autonomous Driving Conditional Generative Modeling Autonomous Driving Simulation Autonomous Driving Scenario Generation+1

Open paper AskPDF arXiv

Generative world models for autonomous driving (AD) are of great value in applications such as data augmentation, closed-loop simulation, and safety-critical scenario evaluation. Unlike the widely studied image modality, in this work we explore generative world models for LiDAR data. Existing generation methods for LiDAR predominantly focus on single frame generation or lack the capacity for interactive simulation, while existing prediction approaches require multiple frames of historical input and can only deterministically predict multiple frames at once. Both paradigms fail to support long-horizon interactive generation. To this end, we introduce \textbf{LaGen}, which, to the best of our knowledge is the first autoregressive framework capable of generating long-horizon LiDAR scenes in a frame-by-frame, interactive manner. LaGen is able to take a single-frame input as a starting point and effectively utilize bounding box information as conditions to generate high-fidelity 4D scene. In addition, we introduce a scene decoupling estimation module to enhance the model's interactive generation capability for object-level content, as well as a noise modulation module to mitigate error accumulation during long-horizon generation. We extensively evaluate LaGen's performance in controlled data generation and long-horizon scene generation on the nuScenes dataset. The experimental results demonstrate that LaGen achieves state-of-the-art performance, especially on later frames. The code is publicly available at: https://github.com/szzhou88/LaGen.

Abstract

Explore similar work

LaGen: Towards Autoregressive LiDAR Scene Generation

OmniLiDAR: A Unified Diffusion Framework for Multi-Domain 3D LiDAR Generation

TAP3D: Thermal-Assisted 3D Human Point Clouds