CMX Lunch Seminar

Denoising diffusion models have led to a series of breakthroughs in image and video generation. Rooted in non-equilibrium thermodynamics, these models enable a variety of extensions by lifting them into augmented spaces, encompassing position, momentum, and potentially additional auxiliary variables. This viewpoint gives rise to a "complete recipe" for constructing invertible diffusion processes. In addition, I will show how the dynamical systems perspective gives rise to new integrators that significantly reduce the number of sampling steps needed at test time, particularly for inverse problems. One rich playground for such inverse problems is the domain of climate science. I will end the talk by demonstrating how recent video diffusion models can effectively super-resolve precipitation patterns in a temporally consistent manner, capturing extremes and local statistics.