A key aspect of generative models is the process of transforming noise into an image. This presentation introduces a novel approach that, instead of converting noise into an image in a single step (as in GANS and VAEs), performs this transformation over multiple stages, each involving different types of noise. In doing so, it departs from conventional methods by sequentially transforming noise into an image across several stages. Additionally, after introducing score-based methods, the integration of these methods with diffusion models through stochastic differential equations will be explored.
