This repo contains PyTorch model definitions, training, and sampling code for our paper https://www.arxiv.org/abs/2507.13366:

Leveraging the Spatial Hierarchy: Coarse-to-fine Trajectory Generation via Cascaded Hybrid Diffusion

In this paper, we propose Cardiff, a coarse-to-fine Cascaded hybrid diffusion-based framework for fine-grained and structure-plausible trajectory generation. By leveraging the hierarchical nature of urban mobility, Cardiff decomposes the generation process into two cascaded levels, i.e., discrete road segment-level and continuous fine-grained GPS-level.

The cascaded framework consists of:

• A segment-level trajectory autoencoder to encode discrete road trajectory into latents;
• A coarse-grained segment-level latent diffusion module
• A conditional fine-grained GPS-level continuous diffusion module

First, download and set up the repo:

git clone https://github.com/urban-mobility-generation/Cardiff.git
cd Cardiff

We provide an environment.yml file that can be used to create a Conda environment.

• Pre-trained checkpoints are stored in saved_models

• We provided a jupyter notebook inference_con.ipynb for quick sampling test.

We provide a training script for Cardiff in train.py.

@article{guo2025leveraging,
  title={Leveraging the Spatial Hierarchy: Coarse-to-fine Trajectory Generation via Cascaded Hybrid Diffusion},
  author={Guo, Baoshen and Hong, Zhiqing and Li, Junyi and Wang, Shenhao and Zhao, Jinhua},
  journal={arXiv preprint arXiv:2507.13366},
  year={2025}
}

Please drop me an email ([email protected]) if you have any questions .