OpenLongTail
Generative Scaling of Long-Tail Driving Data · Paper · Project Page · Code
From a single front-camera driving video, OpenLongTail synthesizes the five missing rig cameras (cross-left, cross-right, rear-left, rear-tele, rear-right) at the same timestamps. One model, built on Wan2.1-VACE-1.3B, renders all five cameras in two stages.
Files
| path | content |
|---|---|
trainable.pt | the OpenLongTail weights (LoRA rank 32 + self-attention, camera and pose embeddings), 40k steps; sha256 486e800113d7d7c11c54f0e37e71f593aebc16b8a0c2f77d0c475c1a5a456c7f |
metadata.json | training metadata |
demo/cached/ | the demo clip as a latent cache: a long-tail scene from the held-out set (a narrow city street after snowfall, lined with parked cars) |
assets/ | the result of the one-click demo on that clip |
The Wan2.1-VACE-1.3B base model is downloaded separately from Wan-AI/Wan2.1-VACE-1.3B.
Usage
git clone https://github.com/phai-lab/OpenLongTail.git && cd OpenLongTail
pip install -r requirements.txt && pip install -e .
bash scripts/download.sh # this checkpoint, the base model, the Wan code and the demo clip
bash scripts/demo.sh # one GPU, ~10-15 min on an H200
License
The weights are released under Apache-2.0. The Wan2.1-VACE-1.3B base model and the other upstream models remain under their own licenses.
Citation
@misc{liu2026openlongtailgenerativescalinglongtail,
title={OpenLongTail: Generative Scaling of Long-Tail Driving Data},
author={Lulin Liu and Nuo Chen and Yan Wang and Bangya Liu and Wenyan Cong and Hezhen Hu and Boris Ivanovic and Hao Wang and Ziyao Zeng and Xinyu Gong and Yang Zhou and Zixiang Xiong and Dilin Wang and Zhangyang Wang and Weisong Shi and Ruohan Zhang and Marco Pavone and Zhiwen Fan},
year={2026},
eprint={2607.09655},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2607.09655},
}