R2Flow

• This is the official implementation of our ICRA 2025 paper, "Fast LiDAR Data Generation with Rectified Flows".
• R2Flow is a rectified flow-based LiDAR generative model, which can generate LiDAR data in few steps.
• Currently, only the inference code is available. We will release the training & evaluation code soon.

traj.mp4

Fast LiDAR Data Generation with Rectified Flows
Kazuto Nakashima, Xiaowen Liu, Tomoya Miyawaki, Yumi Iwashita, Ryo Kurazume
ICRA 2025
Project page | arXiv | Demo

Setup

Dependencies

With Docker & NVIDIA Container Toolkit (recommended for reproducibility):

$ git clone https://github.com/kazuto1011/r2flow.git && cd r2flow
$ docker build --tag r2flow .
$ docker run -it --detach --gpus all --volume $(pwd):/home/user/workspace --name r2flow_container r2flow
$ docker attach r2flow_container

Without Docker:

$ git clone https://github.com/kazuto1011/r2flow.git && cd r2flow
$ pip install -r requirements.txt
$ sudo apt install libsparsehash-dev # for torchsparse
$ pip install git+https://github.com/mit-han-lab/torchsparse.git@v2.0.0
$ pip install natten==0.17.1+torch210cu121 --find-links https://shi-labs.com/natten/wheels/

We tested this code on Ubuntu 22.04 + CUDA 12.1 + NVIDIA RTX 6000 Ada GPUs.

Dataset

Please download a KITTI-360 dataset (163 GB) and make a symbolic link to the dataset at $KITTI360_ROOT:

$ ln -sf $KITTI360_ROOT ./r2flow/data/kitti_360/dataset

Please set the environment variable $HF_DATASETS_CACHE to cache the processed dataset (default: ~/.cache/huggingface/datasets).

Quick demo

Unconditional generation using the pre-trained model:

import torch
import torchdiffeq

# Params
nfe = 256
batch_size = 1
device = "cuda"

# Setup a pre-trained model
model, lidar_utils, cfg = torch.hub.load(
    repo_or_dir="kazuto1011/r2flow",
    model="pretrained_r2flow",
    config="r2flow-kitti360-2rf",  # https://github.com/kazuto1011/r2flow/releases/tag/weights
    device=device,
)

# Euler sampling
t = torch.linspace(0, 1, nfe + 1, device=device)
x0 = torch.randn(batch_size, model.in_channels, *model.resolution, device=device)
with torch.inference_mode():
    x1 = torchdiffeq.odeint(func=model, y0=x0, t=t, method="euler")[-1]

# Post-processing
range_image = lidar_utils.restore_metric_depth(x1[:, [0]])  # range in [0, 80]
rflct_image = lidar_utils.denormalize(x1[:, [1]])  # reflectance in [0, 1]
point_cloud = lidar_utils.convert_metric_depth(range_image, format="cartesian")

We also provide a Gradio-based demo on Hugginface spaces.

Training

Coming soon...

Evaluation

Coming soon...

Citation

If you find this code useful for your research, please cite our paper:

@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},
    year      = 2025,
    booktitle = {Proceedings of the International Conference on Robotics and Automation (ICRA)},
    pages     = {}
}

Acknowledgements

• The most part is based on our previous work (kazuto1011/r2dm).
• Rectified Flow is implemented using torchcfm.
• HDiT is based on crowsonkb/k-diffusion.
• JSD and MMD are based on vzyrianov/lidargen.
• FRID, FSVD, and FPVD are based on hancyran/LiDAR-Diffusion