Generative architectural plan drawings for early design decisions: data grounding and additional training for specific use cases

This study shows the development of models for generating architectural plan drawings to support early design decisions in time–and cost-effective housing projects. Plan drawings are communicated through a combination of visual representations and accompanying textual descriptions. These elements play a crucial role in communicating design specifications to project stakeholders. In particular, by exploring the benefits of early decision-making during the planning phase, this paper explores specific use cases and demonstrates how existing data can be processed for additional training in image-generative models. These models have significantly advanced architectural visualization as a synthetic medium for visual representation. The methodology follows a structured approach: (1) evaluating the efficacy of default base models in generating plan drawings to determine the need for additional data grounding and training; (2) defining the scope of research through a theoretical examination of design requirements for specific use cases for training; (3) systematizing and generalizing the additional training process including data grounding (preparing and preprocessing data suitable for specific use cases), optimizing hyperparameters, and training (implementing models capable of generating images with the desired quality); and (4) demonstrating the proposed applications using the additional training model, specifically within the use case of ‘Korean urban residential housing projects’. This methodology aims to improve the efficiency of design communication during the initial stages of architectural design. By doing so, it enables more effective and strategic planning across a wide range of use cases and scenarios.