UniFuture: A 4D Driving World Model for Future Generation and Perception
TLDR
UniFuture is a unified 4D driving world model that jointly generates future RGB and depth sequences, outperforming specialized models on nuScenes and Waymo.
Reasoning
The paper presents a novel approach to jointly model appearance and geometry in a 4D representation, with strong empirical results on real-world datasets. However, the abstract lacks discussion of limitations or comparisons to other world model paradigms, and the claimed 'world model' terminology is used broadly without explicit connection to interactive or RL settings.
Read-first score
Read-first score 72.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.
Field roles
Rank sensitivity
Stability: volatile; rank range: 163.
Keyword Scores
Deep Analysis
Innovations
- Unified 4D driving world model that jointly models future RGB and depth as coupled projections of the same 4D reality
- Dual-Latent Sharing (DLS) scheme to map visual and geometric modalities into a shared spatio-temporal latent space, implicitly entangling texture with structure
- Multi-scale Latent Interaction (MLI) mechanism enforcing bidirectional consistency between geometry and visual synthesis to prevent structural hallucinations and refine geometric estimation
Methodology
The paper proposes a unified framework that treats future RGB images and depth maps as coupled projections of 4D reality. It uses a Dual-Latent Sharing scheme to entangle texture and structure in a shared latent space, and a Multi-scale Latent Interaction mechanism to enforce bidirectional consistency. The model takes a single current frame and forecasts high-fidelity 4D scene sequences (image-depth pairs). Evaluated on nuScenes and Waymo datasets.
Key Results
UniFuture outperforms specialized models in both future generation and geometry perception on the nuScenes and Waymo datasets, demonstrating the efficacy of unified 4D modeling for autonomous driving.