Automated map generation, especially generating maps from natural language descriptions, not only democratizes access to geographic data but also strengthens decision-making, improves communication and allows for customization. However, map generation faces some challenges, such as lowering the professional threshold, improving generation quality, and ensuring geographic consistency. In recent years, large generative models (e.g., text-to-image models) have excelled in the field of image generation. However, since these models are primarily trained on natural image data, they exhibit significant gaps when generating map data with unique layouts and symbols designed. To address this, we propose a text-to-map generation framework based on diffusion model, called MapGenerator. By using a strategy that combines self-instruct and expert refinement, we construct the training dataset MGTrain and the evaluation dataset MGEval, containing 1000 and 100 pairs of maps and their corresponding detailed descriptions, respectively. Based on the training data, we employ a Parameter-Efficient Fine-Tuning (PEFT) strategy to fine-tune a pre-trained general text-to-image model, enhancing its performance in map generation tasks. Experimental results show that MapGenerator achieves the best FID and CLIP Score among all models, and expert evaluations confirm its superior ability to accurately capture geographic objects and spatial relationships described in the text. The study confirms the feasibility and effectiveness of the diffusion model-based text-to-map generation approach, offering new solutions and technical support for text-driven geographic information generation.
We have released the data used for training MapGenerator, along with evaluation data, which are stored in data.zip. This includes the training data MGTrain, and the evaluation data MGEval. If you have any questions, please emaill us yuwh@cug.edu.cn.
MapGenerator and all our publicly available data are intended for research preview and non-commercial use only, subject to the model License of Stable Diffusion 3.5. Please contact us if you find any potential violations. If you have any questions, you can emaill us yuwh@cug.edu.cn.
MapGenerator has been accepted into the CGIS Journal, If you use data and code from MapGenerator, please declare the following references(Will update later.):