Stable diffusion in simple terms. Learn all about stable diffusion

Imagine you had a model that could give you the probability of an input image being a handwritten digit. You could use this model to generate handwritten digits by altering the input. You can make each pixel slightly lighter or slightly darker and see how that affects the output probability. So all you need is a model that tells you how to alter the input to generate a good output. However, how do you train a model like this?