LLM-Cardiff

Trained models for Coarse-to-Fine Trajectory Generation via LLM-Conditioned Cascaded Diffusion Models — the journal extension of "Leveraging the Spatial Hierarchy: Coarse-to-fine Trajectory Generation via Cascaded Hybrid Diffusion" (KDD 2026, arXiv:2507.13366).

Two cascaded diffusion stages over a spatial hierarchy. Stage 1 generates a route — a sequence of road segments — in the latent space of a frozen autoencoder. Stage 2, conditioned on that latent, generates the GPS points along it.

• Code: https://github.com/urban-mobility-generation/LLM-cardiff
• Dataset: seanguo61/LLM-cardiff-porto

Models

directory	city	condition	stage 1	stage 2	last.pt	ae_mae/
porto/	Porto	5-d attributes	DiT, depth 8	DiT, 64 points	210 MB	65 MB
chengdu/	Chengdu	5-d attributes	DiT, depth 8	DiT, 256 points	204 MB	47 MB
singapore-attr/	Singapore	5-d attributes	DiT, depth 12	DiT, 256 points	337 MB	78 MB
singapore-llm-intent/	Singapore	8 LLM intention slots	DiT, depth 12	DiT, 256 points	327 MB	78 MB

Each directory is self-contained.

file	what it is
last.pt	both diffusion stages and the condition encoders, in one checkpoint
ae_mae/model.pt	the frozen autoencoder: road segments → the latent stage 1 diffuses over
ae_mae/ae_config.json	its architecture, vocabulary size, and which feature table it reads
ae_mae/vocab.json	the segment vocabulary, checked against the dataset's road network at load
ae_mae/segment_features_meta.json	the feature dimensions the encoder expects

Architecture is read from the checkpoint, not from a config, so a depth-8 and a depth-12 stage 1 load through the same code path. The autoencoder is not interchangeable: the diffusion model was trained in that encoder's latent space, and a mismatch aborts at load.

Conditions

	content
5-d attributes	departure slot (10-minute bins), trip distance, number of points, origin cell, destination cell
8 intention slots	origin cell, destination cell, destination place type, home cell, departure slot, day type, distance band, age band — each encoded by a frozen sentence encoder

Use

git clone https://github.com/urban-mobility-generation/LLM-cardiff && cd LLM-cardiff
huggingface-cli download seanguo61/LLM-cardiff --local-dir checkpoints
cp paths.yml.example paths.yml

bash scripts/sample.sh porto          # also: chengdu, singapore-attr, singapore-llm

Sampling reads the dataset: the condition comes from held-out real trips, and the constants that turn the model's output back into coordinates live beside them. Only Porto's data is redistributed; the other three need their data obtained separately.

Citation

If you find these models useful in your research, please cite:

@inproceedings{guo2026leveraging,
  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},
  booktitle={Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1},
  pages={359--370},
  year={2026}
}

The Porto model was trained on data derived from the public ECML/PKDD 15 Taxi Trajectory Prediction dataset (CC BY 4.0); please cite it as well.