For Quanta, science journalist Anil Ananthaswamy traces the backstory of AI image generation. In his telling, a probabilistic model (image) proposed in a 2015 paper by Jascha Sohl-Dickstein set the stage for a subsequent denoising technique now used in DALL-E, Stable Diffusion, and other models. The technique reconciles the gap between generated images and training images, helping generative models provide lucid responses to whimsical prompts like “goldfish slurping Coca-Cola on a beach.”