Addressing the Challenges to AI-Based Generation of Visually Consistent Sets of Map Icons with LoRAicon

Abstract Icons serve everyone’s daily life: in physical places like train stations, on almost every interactive user interface, and on many maps. When multiple icons are used on a map, designers strive for visual consistency of the icon set, as this has been shown to improve legibility of the resulting map. However, creating such map icon sets is time consuming and demands design expertise. With the emergence of generative artificial intelligence (AI), research has been carried out into the generation of icons by AI, enabling non-experts to create map icons. The individual results from existing tools such as PictoAI and DALL·E can vary in style and have too many details. To address this challenge, this research introduces the LoRAicon model, a model integrating the low-rank adaptation (LoRA) fine-tuning method with Stable Diffusion XL (SDXL). The method uses fewer parameters for specific applications and can learn from training icons to produce new icons that maintain visual consistency across a set. This study provides a grounded approach for generating map icon sets, particularly in terms of meeting the critical constraints of visual consistency within a set. For the evaluation, we conducted a user study on two user groups. The result shows that users without design backgrounds perceived LoRAicon as more consistent than users with cartographic design experience. Overall, LoRAicon demonstrated a good level of visual consistency across graphical aspects for both user groups. However, some limitations remain, and further research is needed to address them, particularly when compared to designer-created icon sets.