Definition: AI diffusion

(1) AI adoption in the world. Since ChatGPT, AI has spread like wildfire across the globe. See ChatGPT.

(2) The primary AI technology for image generation today. In the training phase, the neural network (image model) is fed huge numbers of image examples with their meaning (this is a cat); however, images are not stored as JPEGs or other image formats. Each image example is changed by adding a random amount of noise over many steps and then discarded. The resulting diffusion model is a network of neuron weights that were continually adjusted from all the images throughout the many training passes. See AI weights.

At image generation, the inference engine generates a canvas of random noise. It uses the network's weights to gradually remove the noise ("denoise") and create an image based on the user's description. The purpose of the noise at both training and inference is to add variety so that the diffusion process always generates a creative image. See image model and Normal Computing.

Turn an Image Into Noise

This example simply shows noise being added to an image. (Image courtesy of Iguazio Ltd.)