Diffusion model: how AI paints pictures out of noise
A diffusion model is a type of generative AI that creates images, video or sound by gradually cleaning up random noise. Most modern image generators run on models like this.
In plain words
Imagine a fogged-up window where you can see nothing. You wipe it bit by bit, and a landscape slowly appears out of the mist. A diffusion model starts from "fog" - random noise - and wipes it step by step. Only it wipes in such a way that exactly what you described in the prompt shows through.
How it works
- During training, the model is shown millions of pictures with noise added bit by bit until they become pure "snow".
- The model learns to do the reverse: remove a little noise at each step.
- To generate, you start with pure random noise.
- Over dozens of steps the model cleans it up, guided by the text description.
- That is why the same phrase gives a new result every time - the starting noise is different.
Where you'll see it
Image and video generators, AI background fill on photos, extending a shot beyond its edges, and "generative fill" in editors and on smartphones. Most services where you type "draw..." use diffusion models or their relatives.
Does it matter when buying
For buying: if you want to generate and edit images on the device, you will need a powerful chip and lots of memory. Otherwise it runs in the cloud. And a key skill: these models struggle with lettering, the number of fingers and precise details - check the result.
Myths
- Myth: The model finds a ready-made picture in a database and tweaks it. Reality: No, it creates a new image from noise every time.
- Myth: A diffusion model understands the physics of the world. Reality: It copies patterns of appearance, so shadows, reflections and fingers often come out wrong.
FAQ
Why is AI bad at writing text in pictures?
Because it learned letters as visual patterns, not as spelling. Newer models do better.
What is a "seed"?
A number that sets the starting noise. The same seed and prompt give a similar result.
Are there diffusion models for video?
Yes, they clean up noise across many frames at once while keeping the motion.