Generative Artificial Intelligence for Planning Scenario Visualization: A Framework and Benchmark for Controllable Street‐Level Image Synthesis

ABSTRACT

The rapid advancement of Generative AI (GenAI) creates new opportunities for spatial planning. However, most existing applications focus on macro‐scale overhead perspectives, leaving street‐level visualization constrained by strict physical and spatial fidelity challenges. We introduce a controllable GenAI‐human interactive framework for producing spatially faithful and realistic street‐level imagery that reflects envisioned planning scenarios. Using Large‐scale solar photovoltaic (LSPV) site planning as a test case, it leverages vision‐language models to integrate human inputs (e.g., textual intent, scribbles) and existing status photos, synthesizing photorealistic outcomes. To support systematic evaluation, we constructed a large‐scale street‐view imagery dataset of before‐and‐after solar panel installation across the United States. We comprehensively evaluated 11 model settings covering five open‐sourced diffusion models and one commercial vision foundation model to test fidelity and realism. Our study illustrates how a GenAI‐human interactive pipeline can support visualizing proposed environmental interventions, and offers a benchmark for conditioned GenAI visualization applications.