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

Problem
This paper addresses the limitations of existing diffusion-based LiDAR generators, which are typically confined to single-domain settings. Such models necessitate separate training for different datasets or sensing conditions, impeding the ability to synthesize LiDAR data in a unified manner across diverse environmental and operational contexts. The authors present OmniLiDAR, a unified framework capable of generating LiDAR scans across eight distinct domains, including variations in adverse weather, sensor configurations, and acquisition platforms (vehicles, drones, quadrupeds). This work is presented as a preprint and has not yet undergone peer review.

Method
OmniLiDAR employs a text-conditioned diffusion framework that generates LiDAR scans in a shared range-image representation. The core technical contributions include:

1. Cross-Domain Training Strategy (CDTS): This strategy mixes domains within each mini-batch during training, allowing the model to learn from heterogeneous data without isolating optimization by domain.
2. Cross-Domain Feature Modeling (CDFM): This component captures directional dependencies along azimuth and elevation axes, reflecting the anisotropic nature of range images.
3. Domain-Adaptive Feature Scaling (DAFS): A lightweight modulation technique that adjusts for structured domain-dependent feature shifts during the denoising process.
   The model is trained on a newly constructed 8-domain dataset, which integrates real-world LiDAR scans with simulated weather conditions and systematic beam reductions, adhering to official data splits.

Results
The authors report significant improvements in generation fidelity across various downstream tasks. Specifically, OmniLiDAR demonstrates enhanced performance in generative data augmentation for LiDAR semantic segmentation and 3D object detection. The model shows consistent robustness against corruptions, particularly in limited-label scenarios. While specific quantitative results are not detailed in the abstract, the authors claim “strong generation fidelity” and “consistent gains” compared to existing baselines, indicating a substantial effect size in practical applications.

Limitations
The authors acknowledge the absence of a public consolidated benchmark for evaluating multi-domain LiDAR generation, which may limit comparative assessments against other state-of-the-art models. Additionally, the reliance on a single model for diverse domains may introduce challenges in capturing the unique characteristics of each domain, potentially affecting the quality of generated outputs in highly specialized scenarios. The paper does not discuss the computational cost associated with training the unified model or the scalability of the proposed methods.

Why it matters
OmniLiDAR’s unified approach to LiDAR generation has significant implications for scalable simulation and synthetic data creation, particularly in applications requiring diverse sensing conditions. By enabling a single model to handle multiple domains, this work facilitates more efficient data generation processes, which can enhance the training of machine learning models in autonomous systems and robotics. The framework’s ability to perform well in limited-label regimes also suggests potential for improving data efficiency in real-world applications, where labeled data is often scarce.

Authors: Youquan Liu, Weidong Yang, Ao Liang, Xiang Xu, Lingdong Kong, Yang Wu, Dekai Zhu, Xin Li et al.
Source: arXiv:2605.13815
URL: https://arxiv.org/abs/2605.13815v1