Explain like I'm five · AI

What is a Diffusion Model?

Image-generating diffusion models, told as a story about a painter who pulls a picture out of a fog-filled paper, bit by bit.

I can talk… but draw??blank paperdraw me a sunset over the seaa talking parrot cannot draw

The guest asks for a picture. A talking parrot cannot draw.It can describe a sunset over the sea in words, but the paper stays blank.

sunsetsea…words: beans in a rowa different joba picture: a million dots at oncepicking them one by one takes all nighta next-bean bird cannot set a million dots at once

Because a picture is not beans in a row — it is a million dots decided all at once.The parrot picks one next bean at a time. Picking dots one by one would take all night. So a different painter is needed.

NOTEsunset over the seapurple skythe parrot tidies the notepainterfoga bit clearedall clearedwipe the fog bit by bit — the note says which picturediffusion is a painter that pulls a picture out of a fog-filled paper, bit by bit

Diffusion is a painter who pulls a picture out of a fog-filled paper, bit by bit.The paper starts as pure fog (noise). Each wipe makes it a little clearer, and the note says which picture to pull out.

Make the note specificcolor, mood, layout too
Draw several, pick onenever bet on one
More steps, sharperbut slower
Same seed, same picturewrite the number down
step 0step 10step 20step 30more steps, sharper — and slower42same seed number = same picture✓draw several, pick onesteps, seed number, how many to draw — three knobs to play with

Fog, blurry, clearer, sharp — the fog lifts step by step.More steps means sharper but slower. The same seed number gives the same picture, and you draw several to pick the best.

mood, color, layout it does wellit learned the feel of the pictures it sawOPFNodd letters · six fingersthe pile it sawit may have learned someone's whole style — copyrightthe painter learned the feel, not the facts — a person checks the exact parts

Mood and color it does well. Letters and finger counts go wrong often.The painter learned the feel of the pictures it saw, not facts. It may have learned someone's whole style, so copyright questions follow. The parrot that sees pictures is a separate story.

In one breath

Diffusion = a painter who wipes the fog away bit by bit to pull a picture out. The note says which picture, and more steps make it sharper. It learned the feel well, but letters and finger counts go wrong often.

A generative model that starts from pure noise and removes it over many steps to form an image. The prompt is the condition and the seed sets the starting noise. It learned the distribution of its training data, so exact text and anatomy are weak, and training-data copyright questions follow.

When grown-ups say it

Diffusion model
The fog-wiping painter. Stable Diffusion, DALL-E, Midjourney are this painter.
Denoising
One wipe of the fog. The whole of what the painter does: remove a little noise, then a little more.
Steps
How many wipes. More is sharper and slower; fewer is faster and blurrier.
Prompt
The note for the painter. The more specific, the closer the picture. → the trainer's note
Seed
The seed number. Sets the starting fog. Same seed and same note give the same picture.
Generation vs understanding
The painter who draws vs the parrot who sees. Different jobs. → the parrot that sees pictures
Inpainting
Redo one part. Cover just the part to fix with fog and wipe it again.
Copyright · training data
The pile of pictures it saw. It learned from other people's pictures. Some styles it can copy closely, which is under debate.
Next story
The Parrot That Writes Code →
Want another term as a picture book? Request →