UniFuture

UniFuture: A 4D Driving World Model for Future Generation and Perception

Official Hugging Face repository for UniFuture.

Model description

UniFuture is a driving world model that jointly models future scene appearance and geometry within a unified framework.
Given the current image as input, it predicts future RGB-depth pairs for future generation and perception.

According to the project description, UniFuture introduces:

• Dual-Latent Sharing for shared latent learning between image and depth sequences
• Multi-scale Latent Interaction for bidirectional refinement between image and depth features across spatial scales

Paper

UniFuture: A 4D Driving World Model for Future Generation and Perception
ICRA 2026

Repository contents

• unifuture.safetensors: pretrained model weights

Task

This repository is intended for research use on driving world modeling, including:

• future scene generation
• future depth prediction
• future perception

Dataset

The paper reports experiments on nuScenes.

Usage

Please refer to the official GitHub repository for training, evaluation, and detailed usage instructions:

https://github.com/dk-liang/UniFuture

License

Apache-2.0

Citation

@inproceedings{liang2026UniFuture,
  title={UniFuture: A 4D Driving World Model for Future Generation and Perception},
  author={Liang, Dingkang and Zhang, Dingyuan and Zhou, Xin and Tu, Sifan and Feng, Tianrui and Li, Xiaofan and Zhang, Yumeng and Du, Mingyang and Tan, Xiao and Bai, Xiang},
  booktitle={IEEE International Conference on Robotics and Automation},
  year={2026}
}