Diffusion models are generative models that create high-quality data by introducing noise into a data distribution in a controlled manner. The process involves two main phases: the forward diffusion phase, where noise is systematically added to a real image to create pure noise, and the reverse diffusion phase, which learns to reconstruct the original image from the noise. This method, inspired by non-equilibrium statistical physics, effectively breaks down the structure of the data and then restores it, resulting in a flexible generative model capable of producing realistic outputs.
The essential idea is to systematically and slowly destroy structure in a data distribution through an iterative forward diffusion process, then learn a reverse process that restores it.
The diffusion process is split into a forward phase, where we add noise to real images, and a reverse phase, which generates high-quality images from noise.
Collection
[
|
...
]