Generative AI & the senses
A diffusion model is a generative AI that creates images or other data by starting from random noise and removing it step by step until a clear result appears.
Imagine a photo slowly dissolving into TV static. A diffusion model learns that process in reverse. During training, it watches clean images get progressively noisier and learns to predict, at each stage, what noise was added. Once it can do that, it can start from pure static and work backwards to a brand-new image.
Generation happens in a series of small steps, often a few dozen. At each step the model cleans the picture a little, guided by your text prompt so the emerging image matches the description. Each run starts from different random noise, which is why the same prompt gives different pictures.
Diffusion models were first described in 2015 and became practical around 2020. They replaced GANs as the main approach because they are more stable to train and produce more varied results. They now generate video and audio as well as images.
Watching a diffusion model work, you see a blurry gray blob turn into a rough shape, then a recognizable bicycle, then a crisp photo of a red bicycle by a wall.