CityGen: Structure-Guided City-Style Synthesis for Cross-City Autonomous Driving

Researchers introduce CityGen, a diffusion-based framework that enables autonomous driving systems to generalize across different cities without labeled training data. The approach uses HD-map guidance and visual prompts to synthesize city-specific driving scenarios, addressing a critical scalability challenge in deploying autonomous vehicles to new geographic regions.

Autonomous driving deployment faces a fundamental challenge: models trained in one city often fail dramatically when deployed elsewhere due to differences in road layouts, visual appearance, and traffic patterns. This domain shift represents a significant barrier to scaling autonomous vehicle technology globally. CityGen addresses this by proposing a zero-label adaptation framework that doesn't require expensive annotation or city-specific training data, making it more practical for real-world deployment scenarios.

The technical approach leverages generative AI and HD-map conditioning to synthesize realistic driving scenarios tailored to new cities. By using city-level visual prompts rather than individual image labels, the method reduces annotation overhead while maintaining effectiveness across perception, segmentation, and planning tasks. This represents an evolution in how the autonomous driving community thinks about domain adaptation—moving from reactive fine-tuning to proactive synthetic data generation.

For the autonomous driving industry, this work has substantial implications. Current deployment costs are partially driven by the need to collect and annotate large datasets in each target city. A scalable, label-efficient adaptation method could significantly reduce time-to-deployment and operational expenses for autonomous vehicle manufacturers and operators. The introduction of CityTransfer-Bench also provides the community with a standardized evaluation framework for measuring cross-city generalization, which currently lacks standardized benchmarks.

The research establishes foundations for geographically agnostic autonomous systems, potentially accelerating global expansion of autonomous vehicle services. Future work will likely focus on testing this approach in increasingly diverse environments and integrating it into production deployment pipelines.

• →CityGen enables autonomous driving systems to adapt to new cities without labeled target data, reducing deployment costs.
• →The framework uses diffusion-based synthesis guided by HD-maps and city-level visual prompts for zero-label adaptation.
• →CityTransfer-Bench provides the first standardized benchmark for evaluating cross-city generalization in autonomous driving.
• →The approach consistently improves robustness across perception, segmentation, and planning tasks across different cities.
• →Label-efficient adaptation could substantially accelerate global deployment of autonomous vehicle technology.