Optimizing Inference Conditioning Techniques in Image Generation for Participatory Urban Transformation

Generative Artificial Intelligence (GenAI) is emerging as a transformative medium for democratizing participatory urban design, potentially bridging the gap between citizens' conceptualizations and professional representations. While current GenAI tools are divided between professional-grade platforms and accessible solutions for non-experts, technical challenges persist in generating precise and contextually relevant visualizations. This research investigates the optimization of inference conditioning techniques through an open-source approach, implementing an integrated framework within ComfyUI that leverages local computing resources. The methodology combines three key enhancements: prompt engineering through Large Language Models (LLMs), fine-tuning through Low-Rank Adaptation (LoRA), and structural control through ControlNet implementations. Testing this framework on two case studies in Pisa's historical center and suburban area demonstrated how the synergistic combination of these techniques significantly improves the quality and contextual relevance of generated visualizations. Results suggest that advanced conditioning strategies can effectively balance accessibility and precision in participatory urban design tools, supporting the development of more inclusive and sustainable urban transformation processes aligned with the UN's 2030 Agenda goals.