CVD-STORM: Cross-View Video Diffusion with Spatial-Temporal Reconstruction Model for Autonomous Driving

arXiv 25.10 2025 60.6 method, application

TLDR

CVD-STORM generates multi-view driving videos with 4D reconstruction using a cross-view video diffusion model and spatial-temporal VAE.

Reasoning

The paper presents a novel approach combining video diffusion with 4D reconstruction for autonomous driving, showing strong empirical results. However, it lacks interactive or reinforcement learning aspects, limiting its scope to passive video generation.

Read-first score

Read-first score 60.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 41.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=experiment,metric,result

Topical relevance 42%

58.6

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code,github

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 193.

Keyword Scores

generative world model

video world model

world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Cross-view video diffusion model with spatial-temporal reconstruction VAE for autonomous driving
Fine-tuning VAE with auxiliary 4D reconstruction task to enhance encoding of 3D structures and temporal dynamics
Jointly-trained Gaussian Splatting Decoder for dynamic scene reconstruction

Methodology

The approach first fine-tunes a Variational Autoencoder (VAE) with an auxiliary 4D reconstruction task to improve its ability to encode 3D structures and temporal dynamics. This enhanced VAE is then integrated into a video diffusion process to generate long-term, multi-view videos under various control inputs, enabling cross-view video generation with 4D reconstruction capabilities.

Key Results

The model achieves substantial improvements in both FID and FVD metrics, and the jointly-trained Gaussian Splatting Decoder effectively reconstructs dynamic scenes, providing valuable geometric information for scene understanding.

Resources

Paper