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

Simulation Route (a) Initial Scene Generation (b) Long Simulation Rollout (c) Final Destination Figure 1. Overview. SceneDiffuser++ is a single unified, end-to-end trained generative world model that enables CitySim: city-scale traffic simulation that takes in a large map region, start and end points, and simulates everything in between, from initial scene generation, agent behavior prediction, occlusion reasoning, dynamic agent generation (spawning and removal) to environment simulation (traffic lights).