Diffusion models are behind Stable Diffusion, DALLยทE, and Midjourney. The name itself is a clue.
The forward process: adding noise
Start with a real image. Now add a tiny amount of random noise. Then more. After ~1000 steps, you have pure static.
The reverse process: learning to denoise
We train a neural network to undo these noise steps. Given a noisy image, the network learns to predict the original signal.
Generating images
To generate a new image, start with pure noise. Run the denoising network in reverse โ step by step โ and it gradually sculpts a coherent image from the static.