To address the limited real-time performance of existing radio map construction methods in measurement-free scenarios, this work proposes an air-ground collaborative method for shared radio map generation. A diffusion model is introduced to formulate a conditional generation task, thereby meeting the requirement for real-time radio map generation among large-scale distributed user nodes. First, an edge computing architecture enabling shared denoising between Unmanned Aerial Vehicles (UAVs) and ground base stations is constructed. The radio map generation task is formulated as a mixed-integer nonlinear programming problem with the objective of minimizing the total generation latency for all users. Subsequently, the original problem is decomposed into two subproblems: the joint optimization of user-server service policies and shared denoising steps, and UAV trajectory planning. These two subproblems are efficiently solved via the branch-and-bound method and the model-based reinforcement learning algorithm RGCN-DRL, respectively. In particular, a prediction module based on relational graph convolutional networks is designed to effectively capture user mobility intentions, which assists in UAV flight decision-making. Simulation results demonstrate that compared with the baseline method of local radio map generation, the proposed method reduces radio map generation latency by 53.08% while guaranteeing the quality of radio map generation.