Diffusion Model

What is a diffusion model?

You've probably seen AI draw: type a sentence and a few seconds later an image appears. The engine behind most of that is the diffusion model. Its approach is a little counterintuitive — it doesn't "draw" directly; it dirties the canvas, then cleans it up bit by bit.

How does it work?

Forward: add noise
During training, take a real photo and keep adding random noise step by step until it's a field of static.
Reverse: remove noise
Then train the model to go the other way: start from that static and peel the noise away step by step until a clear photo returns. What the model learns is how to "wipe".
Generation: start from noise
Once it knows how to denoise, give it pure noise and it can "wipe" out a brand-new image. Add a text prompt and it wipes in the direction you described.

Why is it everywhere?

Quality
Compared to earlier generative models, diffusion images are richer in detail and more stable, less prone to falling apart.
Control
Text, reference images and local edits can all steer it — it goes where you point.
Open ecosystem
Open-sourcing models like Stable Diffusion spawned a flood of plugins and techniques.

What are its limits?

Generating one image takes many iterative steps, so it's slower and heavier than other models. And since it learns from the whole internet, copyright and bias remain live debates.

Bottom line: a diffusion model "wipes" a picture out of a blob of noise, one step at a time.