TLDR: This research introduces a multi-agent AI system that automates and enhances bicycle infrastructure design on real-world street-view images. It addresses limitations of traditional and existing AI methods by using specialized agents for lane localization, prompt optimization, design generation, and automated evaluation, consistently producing realistic, contextually appropriate, and instruction-compliant street designs for urban planning.
Designing urban spaces, especially for active transportation like cycling, traditionally involves extensive manual effort. Creating realistic visual renderings of proposed street designs is crucial for public engagement and collaborative decision-making. However, these conventional methods are often time-consuming, require specialized graphic design expertise, and make it challenging to quickly adjust designs based on feedback.
While artificial intelligence, particularly generative AI (GenAI), offers a promising path to rapidly create design scenarios, existing approaches often demand large amounts of specific training data and struggle with precise spatial variations in complex street scenes. This is where a new multi-agent system steps in, aiming to transform how bicycle facilities are designed directly onto real-world street-view imagery.
A Multi-Agent Approach to Street Design
Researchers have introduced an innovative multi-agent system that integrates several specialized AI components to overcome these challenges. This framework streamlines the process of editing and redesigning bicycle infrastructure, ensuring the generated designs are realistic, contextually appropriate, and compliant with planning guidelines. The system is built upon a state-of-the-art image generation model, GPT-image-1, and comprises four key agents:
The Locator Agent uses advanced Multimodal Large Language Models (MLLMs) to accurately identify and describe the precise location and features of existing or potential bike lanes within a street-view image. This helps the image generation model understand spatial relationships, a common weakness in standalone image generators.
Next, the Prompt Optimization Agent refines user instructions. It takes a user’s initial prompt and enhances it by incorporating detailed contextual descriptions from the Locator Agent and illustrative examples. This process ensures the image generation model receives clear, unambiguous instructions, significantly reducing semantic misinterpretation and improving output quality.
The Design Generation Agent then takes these optimized prompts and employs a two-step cascading strategy. First, it highlights the target area for the bike lane. Second, it applies the refined prompt to generate multiple candidate designs, such as standard marked lanes, buffered lanes, or colored-surface lanes. This approach helps in robustly rendering complex design elements.
Finally, the Evaluation Agent plays a critical role in selecting the best design. It first filters candidates by comparing them to a reference layout using CLIP similarity, then performs a binary compliance check with reasoning MLLMs to ensure the design adheres to all specified requirements. This two-stage evaluation guarantees that the final selected design is both visually appealing and instruction-aligned.
Also Read:
- MapAgent: A Hierarchical AI Framework for Advanced Geospatial Reasoning
- MSRFormer: A New Approach to Understanding Road Networks Through Multi-scale Interactions
Real-World Application and Impact
Experiments conducted across diverse urban environments, including dense city streets, suburban roads, and complex intersections, demonstrate the system’s adaptability. The qualitative results show that the pipeline consistently embeds various bike lane patterns into different street-view contexts while maintaining correct spatial alignment and overall scene realism. Even in challenging conditions like partial occlusions, the generated lanes remain visually distinct and contextually appropriate.
Quantitative evaluations further support the system’s effectiveness, with the Evaluator Agent achieving over 95% accuracy in correctly selecting suitable candidate designs across eight predefined scenarios. This high level of accuracy underscores the robustness and reliability of the method.
The research also highlights the superior performance of GPT-image-1 as the generation backbone compared to other models like Stable Diffusion 3.5 variants, which often struggled with topological errors, metric inconsistencies, and unintended scene alterations. An ablation study confirmed the crucial contribution of each agent to the overall accuracy and coherence of the generated designs.
This multi-agent framework represents a significant step forward in applying generative AI to urban planning. By decomposing the complex design process into manageable steps handled by specialized agents, the system offers a powerful tool to support bicycle infrastructure planning, facilitate design iterations, and enhance public engagement. For more in-depth information, you can read the full research paper: From Image Generation to Infrastructure Design: a Multi-agent Pipeline for Street Design Generation.
