南方建筑Issue(6)：98-106,9.DOI:10.3969/j.issn.1000-0232.2026.06.009
主观感知驱动的校园空间人工智能生成设计
Artificial Intelligence-generated Design of Campus Space Driven by Subjective Visual Perception
摘要
Abstract
[Purpose]Urban and campus space design proposes an urgent requirement for the accurate translation of users'subjective perceptions into actionable design parameters.However,conventional approaches rely heavily on experience and qualitative judgments,lacking a closed-loop feedback mechanism that quantifies subjective perceptual quantification to design generation.This study explores a design generation pathway based on individual visual perception,machine learning analysis,and generative artificial intelligence.In other words,the nonlinear relationship models between multidimensional perception and spatial elements were built from subjective perceptual feedback data,thereby driving artificial intelligence to generate perceptually targeted scene design solutions. [Methods]A case study based on a university campus was carried out.In the study area,21 panoramic images of typical settings,such as teaching buildings,pathways,plazas,green spaces,and bicycle parking areas—were captured on site.Thirty-four students wore an HTC VIVE Pro Eye headset with integrated eye tracking,browsed the scenes immersively in virtual reality,and their eye fixation distributions were recorded.For each scene,subjective scores for perceived safety,comfort,restorativeness,learning efficiency,and historical sense of respondents were evaluated.All images were semantically segmented to quantify the image proportions of the spatial elements,including vegetation,water bodies,buildings,roads,and bicycles.Gradient boosting decision tree models were built using proportions of visual-spatial elements and eye-tracking heatmap indicators as inputs and five perceptual scores as the outputs.The importance of sequence,nonlinear thresholds,and the interaction effects among different elements were identified.Based on the modeling results,key elements and their critical thresholds were translated into design prompts.Later,the original scene images were re-plotted locally using Stable Diffusion,LoRA fine-tuning,and ControlNet structural guidance,generating an optimized solution.Hence,a complete workflow from the collection of perceptual data,element analysis and modeling,to artificial intelligence generation was established. [Results]Feature importance analysis showed that greenery and water bodies contributed positively to restorativeness;bicycle distribution and the proportion of landscape structures affected safety significantly;sky openness and bicycle ratios were closely related to comfort;and cultural symbols such as sculptures and signage play significant roles in shaping historical sense.The research revealed the general nonlinear threshold effects.For instance,the comfort ratings exhibited a declining inflection point when the proportion of bicycles exceeded a certain value,whereas excessively high ground proportions exerted oppressive negative effects on concentration.Based on the identified thresholds,differentiated design instructions were formulated for two typical scenes.The AI-generated images maintained the original scene structure while also optimizing the visible proportions of natural elements and the distribution of traffic-related components.After multi-criteria quality assessment,the design schemes with visual coherence and spatial guidance were screened. [Conclusions]The mechanism from subjective perceptions of users to generative design instructions was established,which provided a scientific,quantifiable,and user-oriented technical pathway to improve the perceptual quality of campus spaces.The method can help identify perceptual blind spots and can be extended to urban design scenarios such as slow-traffic system improvement and community regeneration.Nevertheless,existing generative AI models still lack precise control over element proportions;the generated images are better suited for conceptual envisioning and strategy formulation.They require further refinement through parametric modeling.Moreover,caution should be taken when applying the derived thresholds to other campus types or cultural contexts,given the limited sample size and scene diversity.Future research should expand participant groups,increase scene variability,and conduct cross-campus comparative studies to strengthen the external validity and engineering applicability of conclusions.关键词
场景感知/人工智能生成式设计/非线性模型/虚拟现实/眼动追踪Key words
scene perception/artificial intelligence generative design/nonlinear model/virtual reality(VR)/eye-tracking分类
建筑与水利引用本文复制引用
陈非,赵方舟,田雨欣..主观感知驱动的校园空间人工智能生成设计[J].南方建筑,2026,(6):98-106,9.基金项目
国家自然科学基金资助项目(52078325):基于多目标优化的既有高校校园绿色化改造评价与设计方法研究——以京津冀地区为例. （52078325）
