Reverse Diffusion Process

What is Reverse Diffusion Process?

The learned process where a neural network (usually a U-Net) predicts and subtracts the noise at each step, iteratively 'denoising' random static back into a coherent image.

Where did the term "Reverse Diffusion Process" come from?

The 'generative' part of diffusion.

How is "Reverse Diffusion Process" used today?

Where the magic happens in Stable Diffusion.

Related Terms