This paper proposes a framework for probabilistic forecasting of dynamical systems using generative modeling. The authors formulate the forecasting problem as sampling from the conditional distribution of the future system state given its current state. They leverage stochastic interpolants to construct a generative model between an arbitrary base distribution and the target distribution. The framework involves designing a fictitious, non-physical stochastic dynamics that takes the current system state as input and produces a sample from the target conditional distribution in finite time without bias. This process maps a point mass centered at the current state to a probabilistic ensemble of forecasts. The drift coefficient in the stochastic differential equation (SDE) is proven to be non-singular and can be efficiently learned via square loss regression over time-series data. The diffusion coefficient in the SDE can also be adjusted post-training, and a specific choice that minimizes the impact of estimation error results in a Föllmer process. The approach is validated on several complex, high-dimensional forecasting problems, including stochastically forced Navier-Stokes equations and video prediction on the KTH and CLEVRER datasets. The paper highlights the utility and scalability of the proposed method, demonstrating its ability to produce diverse samples consistent with the conditional distribution and outperforming standard conditional generative modeling in video generation tasks.This paper proposes a framework for probabilistic forecasting of dynamical systems using generative modeling. The authors formulate the forecasting problem as sampling from the conditional distribution of the future system state given its current state. They leverage stochastic interpolants to construct a generative model between an arbitrary base distribution and the target distribution. The framework involves designing a fictitious, non-physical stochastic dynamics that takes the current system state as input and produces a sample from the target conditional distribution in finite time without bias. This process maps a point mass centered at the current state to a probabilistic ensemble of forecasts. The drift coefficient in the stochastic differential equation (SDE) is proven to be non-singular and can be efficiently learned via square loss regression over time-series data. The diffusion coefficient in the SDE can also be adjusted post-training, and a specific choice that minimizes the impact of estimation error results in a Föllmer process. The approach is validated on several complex, high-dimensional forecasting problems, including stochastically forced Navier-Stokes equations and video prediction on the KTH and CLEVRER datasets. The paper highlights the utility and scalability of the proposed method, demonstrating its ability to produce diverse samples consistent with the conditional distribution and outperforming standard conditional generative modeling in video generation tasks.