# Vista: A Generalizable Driving World Model with High Fidelity and Versatile Controllability

**Authors:** Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, Hongyang Li  
**arXiv:** [2405.17398](https://arxiv.org/abs/2405.17398)  
**Date:** 2024-05-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2405.17398v5/

## Key points

- Introduction of Vista, a cutting-edge driving world model
- Designed for high fidelity and versatile controllability in predicting outcomes of different actions in autonomous driving scenarios
- Incorporates key ingredients identified through a systematic diagnosis of existing methods
- Offers a wide range of controls through an efficient learning strategy
- Incorporates two novel losses and an effective latent replacement approach for realistic predictions at high resolution
- Demonstrates superior performance compared to existing models after extensive training on multiple datasets
- Can establish a generalizable reward for real-world action evaluation without access to ground truth actions
- Acknowledgment of support from various funding sources and gratitude to contributors for valuable feedback
- Potential future work focusing on computation efficiency, quality maintenance, and training scale

## Summary

The authors introduce Vista, a cutting-edge driving world model that addresses key limitations in existing models. Designed for high fidelity and versatile controllability in predicting outcomes of different actions in autonomous driving scenarios, Vista incorporates key ingredients identified through a systematic diagnosis of existing methods. It also offers a wide range of controls through an efficient learning strategy, from high-level intentions to low-level maneuvers. To achieve realistic predictions at high resolution, the model incorporates two novel losses and an effective latent replacement approach. After extensive training on multiple datasets, Vista demonstrates superior performance compared to existing models and can establish a generalizable reward for real-world action evaluation without access to ground truth actions. The authors acknowledge support from various funding sources and express gratitude to contributors for their valuable feedback. They highlight potential future work focusing on computation efficiency, quality maintenance, and training scale.