SynFlow: Scaling Up LiDAR Scene Flow Estimation
with Synthetic Data

Qingwen Zhang Xiaomeng Zhu Chenhan Jiang Patric Jensfelt

KTH HKUST

* The generation pipeline code and dataset will be made publicly available after peer-review.

Overview

Scaling up LiDAR Scene Flow with Synthetic Data. We present SynFlow, a data generation pipeline leveraging the CARLA simulator to synthesize diverse, perfectly labeled LiDAR scene flow data (center). While real-world datasets are often constrained by high annotation costs and limited scenario diversity, SynFlow provides a scalable source of dense, noise-free supervision for learning robust motion priors. As shown in the results (right), models trained on SynFlow-4k dataset achieve strong zero-shot generalization on real-world benchmarks and significantly outperform in-domain baselines when fine-tuned on a small subset of real data. More details available in the main paper pdf.

Supplementary Material

A.1 Overview of Simulated Town Routes

A.2 SynFlow Dataset: Video Previews

Urban Driving Scenarios

Highway Driving Scenarios

Rural & Suburban Scenarios

B. Qualitative Results: Trained Models

Qualitative Result — TruckScenes comparison — Fig. 1. Qualitative comparison on a TruckScenes sequence. The columns show ground truth, an in-domain model (ΔFlow trained on TruckScenes labels), and our zero-shot model trained only on SynFlow-4k. The yellow circle highlights a turning vehicle, with zoomed-in views shown below. In this example, our **SynFlow-4k** model produces motion patterns that are closer to the ground truth and more coherent under turning dynamics. In contrast, the in-domain model shows less consistent motion on the same object, likely due to the limited amount of labeled data available in TruckScenes (only 20% labeled). This example illustrates that physics-consistent synthetic supervision from SynFlow can transfer effectively to real-world LiDAR scene flow when real annotations are scarce.

BibTeX

@article{zhang2026synflow,
  author    = {Zhang, Qingwen and Zhu, Xiaomeng and Jiang, Chenhan and Jensfelt, Patric},
  title     = {SynFlow: Scaling Up LiDAR Scene Flow Estimation with Synthetic Data},
  journal   = {arXiv preprint arXiv:2604.09411},
  year      = {2026},
}