Urban trajectory generation has emerged as a critical research area amidst the complexities introduced by rapid urbaniza- tion. However, due to factors such as privacy protection and deployment costs, collecting urban trajectory data is a long- term and difficult project. Learning human mobility from ex- isting data and generate trajectory in urban area lacking data becomes a critical problem. To address this problem, this pa- per proposes a Generalizable Trajectory Generation model (GTG). The model consists of three parts: 1) cross-city topo- logical feature extraction, which learns local topological fea- tures across cities; 2) road segment cost prediction, which which learns the cross-city mapping between topological fea- tures and travel costs; 3) travel preference learning, which updates the actual preference of each road segment by iterat- ing the shortest path query and preference update. Preference means a combination of multiple travel cost types. Our model surpasses existing models in multiple city data sets and multi- ple macro and micro metrics. Further experiments show that it is sufficient to replace real data for training downstream tasks. Finally, it was verified that after collecting trajectory data of new cities, the model has the ability to be fine-tuned and improved.