What is Diffusion Model? — one analogy

1The analogy

Hang one thin curtain over a window and the view outside blurs a little. Add a second layer, then a third, and shapes start to melt together. Pile on six or seven and the window turns into a pale wall that reveals nothing about what's on the other side.

A diffusion model does this on purpose. It takes a picture and lays blur over it, layer after layer, until the screen holds nothing recognizable. Because it knows the exact order the layers went on, it can also learn the trick of lifting them off one at a time, in reverse.

Making a picture starts at the opposite end: an empty window with nothing behind it at all. The screen holds nothing but blur, and each layer that comes off lets a little more shape settle into place. By the time the last layer lifts, a picture stands there that never existed before.

2In detail

It practices wrecking things first

Training doesn't start on the side that makes pictures — it starts on the side that breaks them. Take one photo and lay down a faint layer of blur. Lay another on top of that, and another, and after a few hundred layers what's left has nothing recognizable about it.

Every one of those in-between stages becomes a paired question and answer. Show the model a screen with three layers of blur on it and ask it to name exactly what the most recent layer looked like. The answer is always known, because it was recorded the moment that layer went on.

Repeat this exercise hundreds of millions of times and the model can look at any screen at all and point out, "the part that looks like blur here is roughly this much." What it builds first isn't the ability to make pictures — it's the ability to recognize blur.

Making a picture means clearing it in reverse

To make a picture, the model doesn't start from a photo — it starts from a fresh screen scattered with meaningless blur. Ask it which part of this looks like blur, and an answer comes back. Take away a little of what it named, and the screen tidies up just slightly.

Show it the tidied screen and ask the same question again. Repeat this back-and-forth dozens of times and shapes that were never there begin to settle in among the blur. It isn't recovering an old photo — it's building a picture that never existed, because the starting screen has nothing to do with any photo at all.

There's a reason it doesn't clear everything in one move. Asking it to jump straight from a blur-covered screen to a finished picture makes the guess wildly wrong. Clearing a little and looking again lets each step correct however far off the last guess was.

The first blur pattern decides the outcome

The blur on the starting screen is scattered at random, but once it's down, that pattern strongly shapes the outcome. Which spots came out a bit darker changes the very first judgment call, and that call carries forward into every step after it.

So the same request produces a picture with a different layout and different colors if the starting pattern is different. Flip it around, though: leave the starting pattern untouched and change only the request, and the layout stays close while only the details shift. Fine-tuning a result you already like rests on exactly this property.

Words can steer the clearing as it happens

Every time the model points out where the blur is, a description can ride along with it. If the description says "a yellow field," the direction it clears in leans that way. Because the description gets consulted again at every step, its pull isn't a one-time nudge — it keeps acting across dozens of steps in a row.

How closely it follows the description can be turned up or down too. Turn it down and the result stays free but drifts from the request; turn it up and it sticks to the request, though colors can blow out and shapes can stiffen. Writing down what shouldn't appear, so it gets pushed the other way, works on the same principle.

How many steps you take splits quality from time

How many steps the clearing is split into is set by a person. Fewer steps means clearing more at once, which makes the guesses coarser and leaves smudged patches in the result. More steps smooths things out, but each added step means more computing, so the wait grows too.

More steps doesn't keep helping forever. Past a certain point the visible difference nearly disappears and only the wait keeps growing. New clearing techniques keep arriving that reach a similar result in fewer steps.

3More precisely

A diffusion model fixes a schedule for adding noise to data little by little until it becomes pure randomness, then trains a neural network to walk that schedule backward. What the network actually outputs isn't a finished picture — it's an estimate of how much noise is mixed into the current screen — and it moves one step at a time by removing that estimate according to a fixed rule.

The analogy breaks down in one place. Lift a curtain and the view that was always there comes back, but behind a diffusion model's starting screen there is no picture at all. The result is built during the clearing, not revealed by it. And where a curtain simply gets pulled aside by hand, diffusion recalculates, at every step, a fresh guess about what counts as blur and how much of it to remove.

Current approaches don't clear pixels directly on the full-size screen. Instead, they shrink the screen down to a much smaller summary, do the clearing inside that smaller space, and expand it back to full size only at the end. That's a large part of why the computing cost dropped enough to run on an ordinary computer.

4Try it yourself

Diffusion Explainer ailearn.space Watch the screen sharpen one step at a time as it moves through the process

5Common misconceptions

It's easy to think diffusion is a technique for sharpening a blurry photo, but actually it invents a new picture from a screen that never held any picture at all.
It's easy to think more steps always makes things better, but actually past a certain point the difference nearly disappears and only the wait grows.
It's easy to think it's slowly reviving a photo that was stored somewhere, but actually no training photo is stored, and the starting point for the clearing has nothing to do with any photo.

7One-line summary

In shortA diffusion model reverses the practice of wrecking a picture with blur, clearing a blur-covered screen step by step until a new picture appears.

Did this analogy help?

Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02