ARCHITECTURAL GENERATIVE MODEL EVALUATION METHODS: IMAGE QUALITY ASSESSMENT METRICS AND EXPERT-BASED APPROACH

The feasibility of using machine learning methods to generative architectural design solutions has been widely recognized as an effective in enhancing innovation, diversity, and efficiency of solutions. However, in generative design methods, the accuracy and quality of design results often rely on empirical evaluation of expert, which is challenging to evaluate and quantify by unified standards. This paper proposes a comprehensive method for evaluating model performance in architectural design tasks. The evaluation is based on computational criteria (i.e., FID, IS, SIMM indicators) and expert system criteria. The computational metrics will measure the distance, diversity, and similarity between the feature vectors of the real image and the generated image. In contrast, the expert criteria will measure the accuracy, intentionality, and rationality of the layout scheme. This study applies this framework to evaluate three widely used generative models in architectural design: GANs, Diffusion Models, and VAE. The framework also guides the optimization of generative models in architectural applications and assists architects in validating generative outcomes with more efficient workflows.