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

2026-05-28 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

CityGen is a diffusion-based generative framework designed to address the performance degradation of autonomous driving systems when deployed in new cities due to significant domain shifts in appearance, road topology, and traffic patterns. It performs zero-label city adaptation through HD-map-conditioned synthesis guided by city-level visual prompts, offering a scalable and label-efficient solution. The framework consistently improves cross-city robustness across perception, segmentation, and planning tasks. Alongside CityGen, the authors introduce CityTransfer-Bench, a geographically disjoint benchmark specifically for evaluating cross-city generalization in autonomous driving, providing a foundation for more generalizable systems.

Key takeaway

For Autonomous Driving Engineers deploying systems in new geographic regions, CityGen offers a critical solution to overcome domain shift challenges without extensive re-labeling. You should investigate integrating diffusion-based generative adaptation techniques, like CityGen's HD-map-conditioned synthesis, to improve your system's cross-city robustness across perception, segmentation, and planning tasks, significantly reducing deployment costs and time.

Key insights

CityGen enables zero-label cross-city adaptation for autonomous driving via HD-map-conditioned diffusion synthesis, improving robustness.

Principles

• Domain shifts severely degrade cross-city autonomous driving performance.
• Zero-label adaptation is crucial for scalable generalization.
• HD-map guidance enhances generative city-style synthesis.

Method

CityGen is a diffusion-based generative framework that synthesizes city styles. It uses HD-map conditioning and city-level visual prompts to achieve zero-label adaptation for autonomous driving tasks.

In practice

• Improve cross-city perception robustness.
• Enhance segmentation performance in new urban areas.
• Strengthen planning capabilities across diverse cities.

Topics

• Autonomous Driving
• Domain Adaptation
• Diffusion Models
• Generative AI
• HD Maps
• CityGen
• CityTransfer-Bench

Best for: Computer Vision Engineer, Research Scientist, Robotics Engineer, AI Scientist, Machine Learning Engineer

Related on AIssential

• DriveCtrl: Conditioned Sim-to-Real Driving Video Generation — DriveCtrl bridges the sim-to-real gap for autonomous driving by generating realistic, temporally consistent, and annotation-preserving videos.
• Using generative AI to improve robots' jumping and landing abilities — Researchers from MIT CSAIL, including Byungchul Kim, Tsun-Hsuan Wang, and Daniela Rus, are exploring the application of generative AI to enhance robotic capabilities.
• RAD-2: Scaling Reinforcement Learning in a Generator-Discriminator Framework — RAD-2 combines a diffusion generator with an RL discriminator for robust, multimodal autonomous driving planning.
• From Seeing to Simulating: Generative High-Fidelity Simulation with Digital Cousins for Generalizable Robot Learning and Evaluation — Digital Cousins generates diverse, high-fidelity simulations from real-world panoramas for robust robot learning.
• OmniLiDAR: A Unified Diffusion Framework for Multi-Domain 3D LiDAR Generation — OmniLiDAR unifies 3D LiDAR generation across diverse sensing domains using a text-conditioned diffusion framework.
• Stage-wise Distortion-Perception Traversal in Zero-shot Inverse Problems with Diffusion Models — A stage-wise diffusion model framework enables flexible distortion-perception tradeoff traversal in zero-shot inverse problems.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.