Vista: A Generalizable Driving World Model with High Fidelity and Versatile Controllability
TLDR
Vista is a generalizable driving world model with high fidelity and versatile controllability, outperforming prior methods on multiple datasets.
Reasoning
The paper introduces novel losses and latent replacement for high-fidelity prediction, and a versatile control set for action controllability. Strengths include strong empirical results (55% FID improvement) and generalization across datasets. Weaknesses are not evident from abstract alone, but the abstract is cut off, potentially missing limitations.
Read-first score
Read-first score 77.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 52.
Field roles
Rank sensitivity
Stability: volatile; rank range: 32.
Keyword Scores
Deep Analysis
Innovations
- Two novel losses to promote learning of moving instances and structural information for high-fidelity prediction
- Latent replacement approach to inject historical frames as priors for coherent long-horizon rollouts
- Versatile set of controls from high-level intentions (command, goal point) to low-level maneuvers (trajectory, angle, speed) via efficient learning strategy
Methodology
Vista is built on a systematic diagnosis of existing driving world models. It introduces two novel losses for moving instances and structural information, a latent replacement method for historical frame priors, and a versatile control learning strategy. The model is trained at large scale and evaluated on multiple datasets against baselines including a general-purpose video generator and the best-performing driving world model.
Key Results
Vista outperforms the most advanced general-purpose video generator in over 70% of comparisons and surpasses the best-performing driving world model by 55% in FID and 27% in FVD. Additionally, Vista itself is used to establish a generalizable reward for real-world action evaluation without ground truth actions.