Skip to content
← AI Folks Labs

It does not draw. It removes static.

Image generators do not paint a picture the way you would — starting with a shape and adding detail. They start with a screen full of random noise and take some of it away, forty times over, until something is left.

7 min · Beginner · Playable

First, the idea

Learn to undo one step of damage

Take a real photograph and add a little static. Then a little more. Do it enough times and nothing is left but noise. Now train a model to do the opposite: given a slightly noisy picture, guess what it looked like one step earlier. That is the entire idea.

Training: destroy a real photograph, one step at a time
real
+28%
+52%
+72%
+88%
noise

Read left to right, that is training. Read right to left, that is generation. The model’s only job is to undo one step.

In plain words

Nobody ever teaches it to draw a cat. It only ever learns to make a noisy thing slightly less noisy. Run that ability forty times starting from pure static, and a picture is what falls out.

Stage 01

Watch it emerge

Press play, or drag the step slider yourself — including backwards, to put the noise back. Notice the order things appear in: the big shapes and the overall light come first, fine detail last. That coarse-to-fine order is not a design choice, it is what the maths does.

Step 0 of 40Pure noise

a unicorn on a hill at dusk

Prompt

Seed 7 — the exact static it starts from. Same seed, same prompt, same picture, every time.

A faithful simulation of the process, not a real model — the target image is drawn by hand so the page needs no download. A real denoiser predicts the noise to remove at each step instead of blending toward a known picture.

In plain words

Guidance is how hard the model is pushed toward your prompt. Turn it down and it wanders off into something vague. Turn it up past ten and colours burn out and edges go hard — the picture obeys the words at the cost of looking real. Every image tool has this dial, usually under a friendlier name.

Stage 02

Why the seed matters

The starting static is not random every time you press go — it comes from a number called the seed. Same seed and same prompt gives the same picture, exactly, forever. Change the seed and you get a different picture of the same description. Press “New seed” above and watch the composition shift while the subject stays.

The prompt

Decides what the picture is of. Steers every step toward matching the words.

The seed

Decides which picture of that thing you get. It is the shape of the initial static.

The steps

Decide how carefully. More steps means finer detail, more time and more cost.

What follows from this

Now the odd behaviour makes sense

Why hands come out wrong

Fine detail is decided in the last few steps, on the smallest scale, with the least information left to work from. Hands are small, highly structured and everywhere in the training data — a hard combination.

Why the same prompt gives different pictures

The prompt sets the direction, the seed sets the starting point. Two different piles of static denoise into two different images of the same description.

Why more steps stops helping

Each step removes a little noise. Once there is barely any left, extra steps cost time and change almost nothing. Past roughly fifty, you are usually paying for nothing.

Why it cannot spell

Text in an image is fine structure that must be exactly right to read as correct. The model is matching the visual texture of letters, not writing words — so it produces things that look like text from across the room.

Next

Same idea, different medium

Language models and image models are more alike than they look — both learn to undo a corruption of their training data. The next lab takes a language model apart the same way.