SceneDiffuser++: City-Scale Traffic Simulation via a Generative World Model

CVPR 25 2025 60.5 method, application

TLDR

SceneDiffuser++ is an end-to-end generative world model for city-scale traffic simulation, integrating scene generation, agent behavior, and dynamic scene control.

Reasoning

The paper introduces a novel end-to-end generative world model that unifies multiple simulation tasks under a single loss function, which is a strength. However, the abstract is truncated, leaving evaluation details and limitations unclear, and the claim of being 'first' requires verification against prior work.

Read-first score

Read-first score 60.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.

Citation impact 18%

87.9

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.8790433

Recency 6%

86.7

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

Methodology quality 18%

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

Topical relevance 29%

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

Reproducibility 18%

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

Citation velocity 12%

8.3

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.50

Field roles

FoundationFrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 389.

Keyword Scores

world model

generative world model

world simulator

world dynamics prediction

interactive world model

video world model

model-based reinforcement learning world model

Deep Analysis

Innovations

First end-to-end generative world model trained on a single loss function capable of point A-to-B simulation on a city scale, integrating scene generation, agent behavior modeling, occlusion reasoning, dynamic scene generation, and environment simulation.
Unified approach to multiple simulation technologies that were previously studied separately, including dynamic scene generation and environment simulation which had received less attention.
Augmented version of the Waymo Open Motion Dataset (WOMD) with larger map regions to support trip-level simulation.

Methodology

SceneDiffuser++ is an end-to-end generative world model trained on a single loss function. It uses the Waymo Open Motion Dataset (WOMD) augmented with larger map regions to enable trip-level simulation. The model is designed to simulate a trip from point A to point B by populating the city around an autonomous vehicle and controlling all dynamic agents (vehicles, pedestrians) and traffic light states.

Key Results

The model demonstrates city-scale traffic simulation capability and achieves superior realism under long simulation conditions. Evaluation is performed on the augmented WOMD dataset.

Limitations

The abstract does not explicitly discuss limitations, but the model is trained and evaluated solely on an augmented version of the Waymo Open Motion Dataset, which may limit generalization to other cities, traffic patterns, or sensor configurations.
As the first end-to-end model of its kind, potential edge cases, scalability to larger cities, and real-time performance constraints are not addressed in the abstract.

Resources

Paper