DiVE: Efficient Multi-View Driving Scenes Generation Based on Video Diffusion Transformer

arXiv 25.4 2025 59.3 method, application

TLDR

DriVerse generates multi-view driving videos from a single image and trajectory using explicit trajectory guidance and motion alignment, outperforming prior models on real datasets.

Reasoning

Strengths include novel trajectory tokenization and motion alignment for temporal consistency, achieving strong results on real-world benchmarks with minimal training. Weaknesses are limited scope to driving scenes and lack of interactivity or downstream task evaluation.

Read-first score

Read-first score 59.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.

Recency 8%

86.7

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

Topical relevance 42%

61.4

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

Methodology quality 25%

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

Reproducibility 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 216.

Keyword Scores

video world model

generative world model

world dynamics prediction

world model

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Explicit trajectory guidance in two complementary forms: tokenizing trajectories into textual prompts using a predefined trend vocabulary, and converting 3D trajectories into 2D spatial motion priors
Lightweight motion alignment module to enhance inter-frame consistency of dynamic pixels

Methodology

DriVerse is a generative model based on Video Diffusion Transformer that simulates navigation-driven driving scenes from a single image and a future trajectory. It introduces dual-form trajectory guidance: tokenizing trajectories into textual prompts for language integration and converting 3D trajectories into 2D spatial motion priors for static content control. A lightweight motion alignment module is added to improve temporal coherence of dynamic objects. The model is trained with minimal additional data and evaluated on nuScenes and Waymo datasets.

Key Results

DriVerse outperforms specialized models on future video generation tasks across both the nuScenes and Waymo datasets.

Resources

Paper