Carving an image out of noise
The diffusion model is the generation technique at the center of image AI. Stable Diffusion, Midjourney, and most video generation services use it — the Diffusion in Stable Diffusion is exactly this.
How it works
In training, noise is added to a clean image step by step until nothing remains but noise, and the model observes that process. It then learns the reverse: what happens when one step of noise is removed.
Generation runs that in reverse. Starting from pure noise, the model removes noise gradually in a direction that matches the prompt, and after a few dozen steps an image emerges. The usual analogy is a sculptor carving a figure out of a block of stone.
Why it took over
The previous mainstream technique, GANs, produced high quality but trained unstably. Diffusion models train stably and produce varied, high-quality images, which made them dominant from around 2022 and the technical basis of the image generation boom. They also worked well with conditioning on text prompts, which mattered a great deal.
Where it shows up in practice
When an image generation service offers a step count setting, that is how many noise-removal steps it takes — a diffusion model parameter. It also explains why each generation differs: the starting noise is random every time. Knowing the mechanism makes the settings and the behavior of these tools considerably more intuitive.
The same approach is applied in video and music generation. Carving a work out of noise has become one of the foundational ideas across generative AI, not only for images.