LIVE: Long-horizon Interactive Video World Modeling
Paper • 2602.03747 • Published • 13
Official checkpoints for LIVE: Long-horizon Interactive Video World Modeling, accepted at ICML 2026 (arXiv:2602.03747).
| File | Role | Format | Size |
|---|---|---|---|
nfd-df-re10k.ckpt |
NFD/DF initialization checkpoint trained on RealEstate10K | PyTorch Lightning checkpoint | 10.345 GB |
nfd-tf-re10k.ckpt |
Teacher Forcing baseline trained on RealEstate10K | PyTorch Lightning checkpoint | 10.361 GB |
live-re10k.ckpt |
Final LIVE checkpoint trained with cycle forcing | PyTorch Lightning checkpoint | 10.437 GB |
vae-kl16.ckpt |
KL16 video VAE used by the models | PyTorch checkpoint | 0.266 GB |
The two NFD baselines and final LIVE model are provided as training checkpoints for use with the official LIVE code release. See the code repository documentation for configuration, data preparation, loading, and inference instructions.
@article{huang2026live,
title={LIVE: Long-horizon Interactive Video World Modeling},
author={Huang, Junchao and Ye, Ziyang and Hu, Xinting and He, Tianyu and Zhang, Guiyu and Shi, Shaoshuai and Bian, Jiang and Jiang, Li},
journal={arXiv preprint arXiv:2602.03747},
year={2026}
}
Please follow the license and usage terms in the official code release and the upstream datasets/models used by this project.