Diffusion Model

Noun · AI & Machine Learning · Origin: 2020

Definitions

A generative AI model that learns to create data by reversing a gradual noising process. Starting from pure noise, it iteratively removes noise to produce coherent images (or audio, video, etc.). DALL-E, Stable Diffusion, and Midjourney all use this approach.
In plain English: An AI that creates images by starting with TV static and gradually removing the noise until a clear picture emerges — like sculpting by removing marble.
The training process adds Gaussian noise to images in T steps until they become pure noise, then trains a neural network to reverse each step. At inference, the model starts from random noise and iteratively denoises it, guided by a text prompt through a mechanism called classifier-free guidance.
Example: 'Our custom diffusion model was trained on 10 million product photos. It generates photorealistic product mockups from text descriptions in under 5 seconds.'
Source: training mechanism

2015: Sohl-Dickstein et al. publish the foundational paper on diffusion probabilistic models, inspired by thermodynamic processes
2020: Ho et al. publish 'Denoising Diffusion Probabilistic Models,' making diffusion models competitive with GANs for image generation
2022: Stable Diffusion, DALL-E 2, and Midjourney bring diffusion models to millions of users, sparking an AI art revolution