Summary: This paper proposes the Consistency Trajectory Model (CTM) to address the limitations of both score-based diffusion models and consistency models. Conventional diffusion sampling methods, such as DDIM and EDM, rely on iterative numerical integration of ODEs or SDEs, which leads to larger discretization errors when the number of function evaluations (NFE) is reduced. In contrast, distillation-based methods enable fast sampling but do not naturally extend to multi-step sampling and offer limited flexibility in controlling the trade-off between NFE and generation fidelity. CTM learns to directly model trajectories between two time points along the diffusion path, thereby combining the accuracy of score-based sampling with the efficiency of consistency-based generation. By introducing soft consistency matching and auxiliary losses, CTM stabilizes training, mitigates gradient vanishing issues, and improves student learning. As a result, CTM provides a unified framework connecting diffusion models and consistency models, achieving state-of-the-art few-step generation performance.
Summary: This paper proposes the Consistency Trajectory Model (CTM) to address the limitations of both score-based diffusion models and consistency models. Conventional diffusion sampling methods, such as DDIM and EDM, rely on iterative numerical integration of ODEs or SDEs, which leads to larger discretization errors when the number of function evaluations (NFE) is reduced. In contrast, distillation-based methods enable fast sampling but do not naturally extend to multi-step sampling and offer limited flexibility in controlling the trade-off between NFE and generation fidelity. CTM learns to directly model trajectories between two time points along the diffusion path, thereby combining the accuracy of score-based sampling with the efficiency of consistency-based generation. By introducing soft consistency matching and auxiliary losses, CTM stabilizes training, mitigates gradient vanishing issues, and improves student learning. As a result, CTM provides a unified framework connecting diffusion models and consistency models, achieving state-of-the-art few-step generation performance.