SUV: Future Scene Understanding as Video Generation for End-to-End Driving
Official checkpoint for SUV: Future Scene Understanding as Video Generation for End-to-End Driving.
SUV casts future scene understanding as video generation. A shared video expert generates future RGB, semantic segmentation, relative depth, and instance-track streams, while an action expert attends to their latent tokens to predict the ego trajectory. The released checkpoint supports the official NAVSIM v1 and NAVSIM v2 evaluation pipelines.
Download
1. SUV checkpoint
mkdir -p SUV_ckpt
hf download ASH-2046/SUV suv_navsim.pt \
--local-dir SUV_ckpt
2. Official Wan2.2 dependency
hf download Wan-AI/Wan2.2-TI2V-5B \
--local-dir SUV_ckpt/Wan2.2-TI2V-5B
Place the complete official Wan2.2 directory next to suv_navsim.pt. It
provides the VAE used during evaluation and the UMT5 encoder and tokenizer used
to prepare the text-embedding cache:
SUV_ckpt/
βββ suv_navsim.pt
βββ Wan2.2-TI2V-5B/
βββ Wan2.2_VAE.pth
βββ models_t5_umt5-xxl-enc-bf16.pth
βββ diffusion_pytorch_model-*.safetensors
βββ google/umt5-xxl/
Set CKPT_LOCAL_DIR in the evaluation shell script to the downloaded
SUV_ckpt directory. Follow the
official code repository for installation,
text-embedding preparation, data setup, and NAVSIM evaluation commands.
Files
| File | Description | SHA256 |
|---|---|---|
suv_navsim.pt |
SUV video and action expert weights | e9f036eff6f2afd4cc4c154cbbe38f50998e700d4d0a5adc43fe28155d55102a |
Citation
@article{yuan2026suv,
title={SUV: Future Scene Understanding as Video Generation for End-to-End Driving},
author={Yuan, Yibo and Fu, Jiacheng and Zhu, Jiangtong and Li, Yi and Han, Jianhua and Tian, Meng and Liu, Zhuohan and Xiong, Zhiwei and Xu, Hang and Fang, Jianwu and Xue, Jianru},
journal={arXiv preprint arXiv:2608.03084},
year={2026}
}
License
The SUV checkpoint is released under the Apache License 2.0. The source code is released separately under the MIT License. Wan2.2 weights are not redistributed in this repository; download them from the official repository and follow its license and citation requirements. Please also follow the requirements of the NAVSIM/NuPlan ecosystem.
Model tree for ASH-2046/SUV
Base model
Wan-AI/Wan2.2-TI2V-5B