DiffMOT is a real-time diffusion-based multiple object tracker that effectively handles non-linear motion patterns. The tracker uses a novel Decoupled Diffusion-based Motion Predictor (D²MP) to model the motion distribution of objects and predict their future positions. D²MP is designed to generate motion from a normal distribution conditioned on historical motion information, enabling accurate and efficient non-linear motion prediction. DiffMOT achieves high performance on the DanceTrack and SportsMOT datasets, with HOTA metrics of 62.3% and 76.2%, respectively, and operates at 22.7 FPS. The tracker outperforms state-of-the-art methods in non-linear motion tracking and demonstrates robustness in handling complex scenarios. The proposed method introduces a diffusion probabilistic model into MOT for the first time, enabling accurate non-linear motion prediction while maintaining real-time performance. DiffMOT is evaluated on multiple datasets, including MOT17, and shows comparable performance to SOTA methods in pedestrian-dominant scenarios. The method is efficient, with a one-step sampling process that reduces inference time and enables real-time tracking. The results demonstrate that DiffMOT is a strong candidate for real-world applications in multi-object tracking.DiffMOT is a real-time diffusion-based multiple object tracker that effectively handles non-linear motion patterns. The tracker uses a novel Decoupled Diffusion-based Motion Predictor (D²MP) to model the motion distribution of objects and predict their future positions. D²MP is designed to generate motion from a normal distribution conditioned on historical motion information, enabling accurate and efficient non-linear motion prediction. DiffMOT achieves high performance on the DanceTrack and SportsMOT datasets, with HOTA metrics of 62.3% and 76.2%, respectively, and operates at 22.7 FPS. The tracker outperforms state-of-the-art methods in non-linear motion tracking and demonstrates robustness in handling complex scenarios. The proposed method introduces a diffusion probabilistic model into MOT for the first time, enabling accurate non-linear motion prediction while maintaining real-time performance. DiffMOT is evaluated on multiple datasets, including MOT17, and shows comparable performance to SOTA methods in pedestrian-dominant scenarios. The method is efficient, with a one-step sampling process that reduces inference time and enables real-time tracking. The results demonstrate that DiffMOT is a strong candidate for real-world applications in multi-object tracking.