Abstract
The style of graphic design is an important design factor that influences the memorability of designs. Graphic designers routinely analyze the latest designs, capture the style trends, and create designs that match the style trends to appeal to a larger audience. Nonetheless, the lack of quantitative style analysis techniques can lead to an inefficient analysis process and introduce subjectivity and bias. To expedite designers’ understanding of design style trends and make the analysis more objective, we propose GradeS, an AI-driven graphic design support system that facilitates multifaceted quantitative analysis of graphic design style. The system was designed and developed in collaboration with designers and comprises four primary interfaces: GradeS:S, GradeS:Q, GradeS:C, and GradeS:T, each serving specific needs identified through interviews with designers. We leveraged the Vision Transformer to model the one-to-many relationship between designs and styles and implemented all interfaces based on the quantitative style representation learned by the model. To train the model, we built a graphic design dataset with carefully designed coarse-grained style labels. We have released the dataset to the community to promote research in data-driven design. To demonstrate the effectiveness of our study, we evaluated both the model and the system. Our model exhibits superior performance in style classification compared to CLIP. Through a user study involving six designers, our system’s effectiveness in supporting designers in analyzing style quantitatively, capturing style trends comprehensively and quickly, and further stimulating creative thinking was demonstrated.
