ControlTraj is a novel framework for generating high-fidelity, controllable trajectories using a topology-constrained diffusion model. The framework addresses the challenges of trajectory generation by integrating road network topology constraints to guide the geographical outcomes. Key contributions include: 1. **RoadMAE**: A masked road segment autoencoder that captures fine-grained road segment embeddings, enabling the model to learn from road network topology. 2. **GeoUNet**: A geographic attention-based denoising UNet architecture that integrates topological constraints into the diffusion process, allowing for flexible and controllable trajectory generation. 3. **Controlled Trajectory Generation**: The framework can generate trajectories that adhere to specific conditions and constraints, such as road network topology and trip attributes. The method is evaluated on three real-world datasets (Chengdu, Xi'an, and Porto) and demonstrates superior performance in terms of fidelity, flexibility, and generalizability compared to state-of-the-art baselines. The results show that ControlTraj can produce high-fidelity trajectories with competitive utility, and the generated trajectories exhibit strong generalizability to unexplored road network topologies. The framework's ability to control trajectory generation and its adaptability to new urban environments without retraining are highlighted as key advantages.ControlTraj is a novel framework for generating high-fidelity, controllable trajectories using a topology-constrained diffusion model. The framework addresses the challenges of trajectory generation by integrating road network topology constraints to guide the geographical outcomes. Key contributions include: 1. **RoadMAE**: A masked road segment autoencoder that captures fine-grained road segment embeddings, enabling the model to learn from road network topology. 2. **GeoUNet**: A geographic attention-based denoising UNet architecture that integrates topological constraints into the diffusion process, allowing for flexible and controllable trajectory generation. 3. **Controlled Trajectory Generation**: The framework can generate trajectories that adhere to specific conditions and constraints, such as road network topology and trip attributes. The method is evaluated on three real-world datasets (Chengdu, Xi'an, and Porto) and demonstrates superior performance in terms of fidelity, flexibility, and generalizability compared to state-of-the-art baselines. The results show that ControlTraj can produce high-fidelity trajectories with competitive utility, and the generated trajectories exhibit strong generalizability to unexplored road network topologies. The framework's ability to control trajectory generation and its adaptability to new urban environments without retraining are highlighted as key advantages.