High-order Langevin dynamics (HOLD) can simultaneously model position, velocity, and acceleration, thereby improving the quality and speed of data generation in diffusion generative modeling.
