Fast LiDAR Data Generation with Rectified Flows

TL;DR: A straight flow-based generative model for fast LiDAR data generation, named R2Flow

Abstract

Building LiDAR generative models holds promise as powerful data priors for restoration, scene manipulation, and scalable simulation in autonomous mobile robots. In recent years, approaches using diffusion models have emerged, significantly improving training stability and generation quality. Despite their success, diffusion models require numerous iterations of running neural networks to generate high-quality samples, making the increasing computational cost a potential barrier for robotics applications. To address this challenge, this paper presents R2Flow, a fast and high-fidelity generative model for LiDAR data. Our method is based on rectified flows that learn straight trajectories, simulating data generation with significantly fewer sampling steps compared to diffusion models. We also propose an efficient Transformer-based model architecture for processing the image representation of LiDAR range and reflectance measurements. Our experiments on unconditional LiDAR data generation using the KITTI-360 dataset demonstrate the effectiveness of our approach in terms of both efficiency and quality.

Method Overview

What kind of framework is suitable for representing LiDAR data?

Recent advances in LiDAR generative models are driven by diffusion models based on the range image representation. Among them, LiDAR diffusion models currently follow two approaches for generating range images: pixel-space iteration and feature-space iteration, inspired by a natural image domain. We prioritize pixel precision required for range images and employ the pixel-space iteration approach.

	Pixel-space iteration [Ho et al. 2020]	Feature-space iteration [Rombach et al. 2022]
Architecture	Diffusion modeling on high-dimensional pixels	Diffusion modeling on lower-dimensional features compressed by autoencoders (AEs)
Pros	Finer details via direct iterative modeling	The iterative space is low-dimensional; high throughput
Cons	The iterative space is high-dimensional; low throughput	Blurriness due to the extra AE decoding
LiDAR application	LiDARGen [Zyrianov et al. 2022] R2DM [Nakashima et al. 2024] R2Flow (ours)	LiDM [Ran et al. 2024]

How can we reduce the number of sampling steps?

Sampling of diffusion models requires iterative evaluations of neural networks such as U-Net. The overall throughput is highly dependent on how much the number of iterations can be reduced. We address this issue by adopting rectified flows [Liu et al. 2023] which are robust to the number of sampling steps.

	Diffusion models [Song et al. 2021]	Rectified flows [Liu et al. 2023]
Formulation	Stochastic differential equations (SDEs)	Ordinary differential equations (ODEs)
Trajectory	Probabilistic & curved (prone to discretization errors)	Deterministic & straight (easy to approximate with few steps)
LiDAR application	LiDARGen [Zyrianov et al. 2022] R2DM [Nakashima et al. 2024] LiDM [Ran et al. 2024]	R2Flow (ours)

What if the pixel-space iteration is completed in just a few steps?

By combining the pixel-space iteration approach with rectified flows, we demonstrated both high geometric accuracy and generation efficiency. Moreover, we also propose a Vision Transformer-based architecture for the neural network that learns flows in pixel space. We modify HDiT (hourglass diffusion transformer) [Crowson et al. 2024] to represent LiDAR range & reflectance (intensity) images.

Results

Unconditional generation

The following compares LiDAR generation models trained on the KITTI-360 dataset. For the diffusion-based methods (LiDM & R2DM) and R2Flow, we show samples generated with many steps and few steps. R2Flow shows consistent quality over different number of steps. Please see our paper for quantitative results on generation quality and computational costs. Note: LiDARGen runs on the fixed noise schdule. LiDM only produces the range modality.

Citation

@inproceedings{nakashima2025fast,
    title     = {Fast LiDAR Data Generation with Rectified Flows},
    author    = {Kazuto Nakashima and Xiaowen Liu and Tomoya Miyawaki and Yumi Iwashita and Ryo Kurazume},
    booktitle = {Proceedings of the IEEE International Conference on Robotics and Automation (ICRA)},
    pages     = {},
    year      = 2025
}

Acknowledgments

This work was supported by JSPS KAKENHI Grant Number JP23K16974 and JSPS KAKENHI Grant Number JP20H00230