TL;DR InfGen performs interleaved long-term closed-loop motion simulation and scene generation with unified next-token prediction.

Long-term Traffic Simulation with Interleaved Autoregressive Motion and Scenario Generation
Xiuyu Yang*, Shuhan Tan*, Philipp Krähenbühl (* equal contribution)
ICCV 2025 (arXiv 2506.17213)

If you find our work useful in your research, please consider citing our paper:

@inproceedings{yang2025infgen,
    title={Long-term Traffic Simulation with Interleaved Autoregressive Motion and Scenario Generation},
    author={Yang, Xiuyu and Tan, Shuhan and Kr{\"a}henb{\"u}hl, Philipp},
    booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    year={2025}
}

Run the following script, and create new environemnt with name infgen (default)

git --recurse-submodules clone https://github.com/OrangeSodahub/InfGen.git
cd InfGen/

conda env create -f environment.yaml [-n ENV_NAME]
conda activate infgen

Alternatively, you can refer to SMART#requirements for manually installations.

NOTE: We have similar execution steps to SMART, but the outputs of Step3 are not identical.

Step 1: Download the Dataset

Download the Waymo Open Motion Dataset (scenario protocol format) from here. The version of WOMD used in InfGen is v1.2.1.

Step 2: Install the Waymo Open Dataset API

Follow the instructions here to install the Waymo Open Dataset API.

Step 3: Preprocess the Dataset

Preprocess the dataset by running: where $SPLIT is chosen from "training" and "validation".

bash scripts/data_preprocess.sh $SPLIT [--input_dir] [--output_dir]

The first path is the raw data path, and the second is the output data path.

The processed data will be saved to the data/waymo_processed/ directory as follows:

InfGen
├── data
│   ├── waymo_processed
│   │   ├── training
│   │   ├── validation
│   │   ├──testing
├── model
├── utils

Run the following script: where set the number of processes in DDP by $GPUS

bash scripts/run_train.sh $GPUS

First set the CKPT_PATH in the following script and run:

bash scripts/run_eval.sh $GPUS

• Release arXiv technical report
• Release full codes
• Release other detailed instructions

Thansk for these excellent opensource works and models: SMART; CatK.