Intelligence assessment for diffusion-based generation of gymnasium field-level plans using deep learning

Abstract Field-level plan design for large and medium-sized gymnasiums is highly complex and inefficient under traditional manual workflows, motivating the use of generative AI. This study proposes a diffusion-based, stepwise assessment framework for gymnasium plan design. In the first stage, Stable Diffusion fine-tuned with LoRA and constrained by ControlNet is used to generate block plans. A rule-based screening module removes outputs with poor visual quality or missing essential functions, while the CNN-based model further ranks the remaining results by topological similarity to 149 exemplary built cases. Finally, the high-quality alternatives were determined by architects. These assessment phases between generation steps improve semantic and functional alignment with building code and reference cases. In the second stage, these selected block plans are further translated into detailed plans with room-level separations through image-to-image diffusion. The two level evaluation system checks how well the detailed plans match the block plan and architectural standards. Architects then choose the scheme that best fits the design intent. The proposed method was applied to the plan generation of the Beijing Jiaotong University Campus Gymnasium in Xiong’an. The generated plan is comparable to the winning bid implemented plan in terms of functional zoning and spatial organization. By embedding assessment phases between generation steps, the framework forms an integrated, feedback-enabled generative assessment paradigm for human computer interaction. Its multi-level, constraint-aware representation links block and detailed plans, maintaining consistency from functional zoning to room-level layouts while reducing manual screening effort.