Diffusion: Definition and Examples

Diffusion is an artificial intelligence generation paradigm inspired by thermodynamics. The principle is based on two phases: a forward process where Gaussian noise is progressively added to a real data point until it becomes unrecognizable, then a reverse process where a neural network learns to remove this noise step by step to reconstruct coherent data from pure noise. <br><br>Concretely, during training, the model observes thousands of images at various noise levels and learns to predict the noise added at each step. Once trained, it can start from a random noise image and gradually "denoise" it to generate an entirely new image. This iterative process, typically involving between 20 and 1000 steps, is what gives diffusion models their remarkable quality. <br><br>Diffusion models are at the core of the most popular image generation tools like Stable Diffusion, DALL·E, and Midjourney. Their strength lies in training stability (unlike GANs), their ability to produce highly detailed images, and their natural compatibility with text conditioning. This is called text-to-image: a text prompt guides the denoising process to steer generation toward the desired result. <br><br>Beyond images, the diffusion principle now extends to audio generation (music, voice), video, 3D models, and even molecular design. Variants like latent diffusion operate in a compressed space to reduce computational cost, making these models accessible on consumer hardware.