ABSTRACT:Urban spatial structure is a critical and complex component of urban design. This study proposes a method for large model training within the urban planning domain to generate and evaluate the local morphology of high-density urban districts，enabling rapid arrangement and assessment of urban spatial elements through generative design.The study selected images of several high-density areas in three cities, including road networks, buildings, greening and other elements. The ComfyUI platform is employed for Low-Rank Adaptation (LoRA) fine-tuning training to enhance the semantic representation of urban spaces. By integrating geometric constraints of ControlNet, local updates are performed on spatial elements such as road structures, building layouts, and metro station locations. Generated results are quantitatively evaluated using space syntax theory and compared with original road networks through multidimensional analysis.The findings demonstrate that domain-specific learning with large models effectively captures the spatial morphological characteristics of high-density urban districts, integrating urban spatial form design knowledge into general image generation models. The generated results fine tune and optimize the urban land use on the basis of keeping consistent with the visual and structural requirements of the real world urban space.This study validates the application potential of generative models based on large model learning in complex urban spatial layouts, assisting designers in rapid urban renewal design. Keywords: Urban Planning-Specific Large Model, Urban Spatial Form, Large - Model Fine - Tuning, Space Syntax