Reconsidering Design Pedagogy through Diffusion Models

Abstract

The text-to-image based diffusion models are deep learning models that generate images from text-based narratives in user-generated prompts. These models use natural language processing (NLP) techniques to recognize narratives and generate corresponding images. This study associates the assignment-based learning-by-doing of design studio with the prompt-based diffusion models that require fine-tuning in each image generation. The reference is a specific formal education setup developed within the context of compulsory courses in design programs' curricula. We explore the implications of diffusion models for a model of the basic design studio as a case study. The term basic design implies a core and foundational element of design. To explore and evaluate the potential of AI tools to improve novice designers' design problem solving capabilities, a retrospective analysis was conducted for a series of basic design studio assignments. The first step of the study was to reframe the assignment briefs as design problems and student design works as design solutions. The outcomes of the identification were further used as input data to generate synthetic design solutions by text-to-image diffusion models. In the third step, the design solution sets generated by students and the diffusion models were comparatively assessed by design experts with regards to how well they answered to the design problems defined in the briefs. The initial findings showed that diffusion models were able to generate a myriad of design solutions in a short time. It is conjectured that this might help students to easily understand the ill-defined design problem requirements and generate visual concepts based on written descriptions. However, the comparison indicated the value of design reasoning conveyed in the studio, as it gets highlighted with the lack of improvement in the learning curve of the diffusion model recorded through the synthetic design process.