The core generative mechanism in Diffusion Models. While the 'forward process' gradually destroys an image by adding Gaussian noise until it becomes pure random static, the 'reverse process' trains a neural network (typically a U-Net) to predict and subtract that noise step-by-step. By iteratively applying this denoising operation, the model can generate coherent, high-quality images starting from pure random noise.
Fundamental concept in Diffusion Probabilistic Models (Ho et al., 2020).
The mechanism powering Stable Diffusion, DALL-E 2, and Midjourney.