LiDAR-GS++:Improving LiDAR Gaussian Reconstruction via Diffusion Priors

TL;DR

LiDAR-GS++ tackles the problem of extrapolation artifacts in GS-based LiDAR reconstruction for open-road re-simulation. It introduces a diffusion-prior–assisted LiDAR reconstruction framework that combines a neural 2DGS field with a controllable LiDAR generation model conditioned on coarsely extrapolated rendering, enabling geometry-consistent expansion beyond observed viewpoints. A depth distortion-aware distillation strategy selectively refines under-fitted regions to harmonize generated data with real scans, achieving state-of-the-art performance on both interpolated and extrapolated viewpoints across public datasets. The approach offers real-time rendering with high fidelity and demonstrates practical impact for scalable, multi-view LiDAR re-simulation in autonomous driving research and testing.

Abstract

Recent GS-based rendering has made significant progress for LiDAR, surpassing Neural Radiance Fields (NeRF) in both quality and speed. However, these methods exhibit artifacts in extrapolated novel view synthesis due to the incomplete reconstruction from single traversal scans. To address this limitation, we present LiDAR-GS++, a LiDAR Gaussian Splatting reconstruction method enhanced by diffusion priors for real-time and high-fidelity re-simulation on public urban roads. Specifically, we introduce a controllable LiDAR generation model conditioned on coarsely extrapolated rendering to produce extra geometry-consistent scans and employ an effective distillation mechanism for expansive reconstruction. By extending reconstruction to under-fitted regions, our approach ensures global geometric consistency for extrapolative novel views while preserving detailed scene surfaces captured by sensors. Experiments on multiple public datasets demonstrate that LiDAR-GS++ achieves state-of-the-art performance for both interpolated and extrapolated viewpoints, surpassing existing GS and NeRF-based methods.

Figures (7)

• Figure 1: LiDAR re-simulation methods, such as GS-LiDAR gsldiar and LiDAR-GS lidargs, encounter performance drops when rendering extrapolated views, e.g., lateral viewpoint shifting. In contrast, LiDAR-GS++ maintains stable and reliable performance by diffusing priors.
• Figure 2: The workflow of LiDAR-GS++. Given an input driving scan, we first project the point cloud to the range view and reconstruct the scene using a neural 2DGS field from a single traverse. Then, we feed coarsely extrapolated rendering to a pre-trained LiDAR diffusion model to produce geometrically consistent extrapolated scans as extra supervision signals, and utilize a depth distortion-aware strategy to distill the GS representation for expansive reconstruction, where the distorted areas of the extrapolated view are highlighted in red. During inference, rendering from reconstructed Gaussians enables real-time, high-quality synthesis for both extrapolated and interpolated viewpoints. Subsequently, the range view can be converted into a point cloud format via inverse projection.
• Figure 3: Design and workflow of the Neural 2DGS Field. The numbers on the vector represent the dimensions.
• Figure 4: Workflow of the Controllable LiDAR Generation.
• Figure 5: Qualitative and quantitative comparisons of our coarse-to-fine cross-lane novel view synthesis on the Para-Lane dataset. (a) Real LiDAR scan as a reference from another-pass; (b) Rendered without diffusion priors, and the red point indicates the distorted area $M$; (c) Generative result of controllable LiDAR diffusion model conditioned by coarsely extrapolated rendering; (d) Rendering after expansive reconstruction with diffusion priors.