ControlTraj is a controllable trajectory generation framework that uses a topology-constrained diffusion model to generate high-fidelity, adaptable trajectories. Unlike previous methods, ControlTraj integrates road network topology to guide trajectory generation, ensuring that generated paths align with real-world geography. The framework includes a novel road segment autoencoder (RoadMAE) that captures fine-grained road embeddings and a geographic attention-based UNet (GeoUNet) that synthesizes trajectories from noise. By combining these components, ControlTraj can generate realistic, controllable trajectories that adapt to new environments without retraining. ControlTraj addresses the limitations of existing trajectory generation methods by ensuring high fidelity, flexibility, and generalizability. It is evaluated on three real-world datasets and demonstrates superior performance in terms of spatial-temporal accuracy, realism, and adaptability. The model's ability to generate trajectories that conform to road network constraints and human-defined preferences makes it suitable for applications such as urban planning and navigation systems. Additionally, ControlTraj shows strong generalizability across different geographical contexts, as demonstrated by zero-shot learning experiments where it outperforms other models in adapting to new cities. The framework's key contributions include the development of RoadMAE for capturing detailed road embeddings and GeoUNet for integrating topological constraints into the diffusion process. These components enable the model to generate high-fidelity trajectories that are both realistic and controllable. The results highlight the potential of ControlTraj in simulating realistic human mobility patterns under various constraints, making it a promising solution for trajectory generation in real-world applications.ControlTraj is a controllable trajectory generation framework that uses a topology-constrained diffusion model to generate high-fidelity, adaptable trajectories. Unlike previous methods, ControlTraj integrates road network topology to guide trajectory generation, ensuring that generated paths align with real-world geography. The framework includes a novel road segment autoencoder (RoadMAE) that captures fine-grained road embeddings and a geographic attention-based UNet (GeoUNet) that synthesizes trajectories from noise. By combining these components, ControlTraj can generate realistic, controllable trajectories that adapt to new environments without retraining. ControlTraj addresses the limitations of existing trajectory generation methods by ensuring high fidelity, flexibility, and generalizability. It is evaluated on three real-world datasets and demonstrates superior performance in terms of spatial-temporal accuracy, realism, and adaptability. The model's ability to generate trajectories that conform to road network constraints and human-defined preferences makes it suitable for applications such as urban planning and navigation systems. Additionally, ControlTraj shows strong generalizability across different geographical contexts, as demonstrated by zero-shot learning experiments where it outperforms other models in adapting to new cities. The framework's key contributions include the development of RoadMAE for capturing detailed road embeddings and GeoUNet for integrating topological constraints into the diffusion process. These components enable the model to generate high-fidelity trajectories that are both realistic and controllable. The results highlight the potential of ControlTraj in simulating realistic human mobility patterns under various constraints, making it a promising solution for trajectory generation in real-world applications.