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.
The 'generative' part of diffusion.
Where the magic happens in Stable Diffusion.