This paper presents a new generative modeling framework that is based on phase space dynamics, drawing inspiration from Critically Damped Langevin Dynamics (CLD). By incorporating ideas from stochastic optimal control, we develop a path measure in the phase space that is highly effective for generative sampling. One key aspect of our approach is its ability to predict data early on by using propagating Ordinary Differential Equations (ODEs) or Stochastic Differential Equations (SDEs) processes. This early prediction, made possible by the unique structure of our model, paves the way for more efficient data generation by utilizing velocity information along the trajectory. This innovation has led to a new method of reducing sampling complexity by going directly from noisy data to authentic images.
Our model produces results in image generation that are comparable to existing methods, and excels particularly in scenarios with a limited Number of Function Evaluations (NFE). Additionally, our approach performs on par with diffusion models that have efficient sampling techniques, highlighting its potential in the field of generative modeling.
