Diffusion: images from noise

What this lesson shows

The model starts from pure noise and removes a little at each step; it was trained to denoise, and your prompt, embedded in the same space you already know, steers every step toward a matching image.

The question it opens with

You type a sentence and get a photo that never existed. Predicting the next word explains none of it, so where did the image come from?

The walkthrough, in the lesson's own words

• A photo appears that never existed. Where did it come from? Make a guess.
• Drag the noise out. Watch a picture surface from static.
• Switch the prompt. The same denoiser, steered toward a new subject.
• Image generation is trained denoising, guided by your prompt's vector.
• This is the engine inside Midjourney, DALL-E, and every image generator you have tried.
• You typed a sentence and got a photo. How did the model build it?
• Not painted. The model never lays down one final pixel at a time; there is no brush and no plan of the finished image.
• Not retrieved. The photo never existed anywhere. The model starts from random static and removes noise until an image is left.
• Exactly. It starts from pure random noise and removes a little at each step. It was trained to denoise, so an image is what remains.
• During training the model saw millions of real images with noise added, and learned to predict the clean version. Point it at pure static and it does the same thing: it removes what looks like noise, step by step, and a plausible image is what is left. Different starting static gives a different image, which is why the photo never existed before.
• A toy image on a tiny grid. A real denoiser walks over millions of pixels across dozens of steps, not a handful of dots.
• From static to a picture, one step of denoising at a time.
• At full noise it is random static. Pull the noise out.
• That is the whole trick. The model starts from noise and removes a little at each step, and a picture is what is left. This is why it is called a diffusion model.
• Same denoiser, same noise. Your prompt is turned into a vector, the same kind of embedding you already met, and it steers every step toward a matching image.
• Two toy prompts. A real prompt is a rich vector that nudges thousands of denoising steps, not a switch between two shapes.
• The prompt vector for a cat pulls each step toward cat-shaped structure. Change the vector and the same denoiser lands somewhere else.
• Now the vector says sun, so the identical denoiser walks the same noise toward a sun instead. The prompt is the steering wheel, not the engine.
• At every step the model asks two questions: what does clean look like on its own, and what does clean look like given this prompt vector. It then leans toward the prompt's answer. Turn that lean up and the image matches the prompt more tightly; turn it down and it drifts back to a generic image. That lean is called guidance.
• You watched noise turn into a picture step by step, and steered it with a prompt. Image generation is trained denoising, guided by your prompt's vector.
• An image generator predicts the next pixel, the way a chatbot predicts the next word.
• It is a denoiser, not a next-token predictor. It starts from pure noise and removes a little at each step, and your prompt's vector steers every step toward a matching image.
• A friend says AI images are just the model picking one pixel after another, like autocomplete for pictures. Using this lesson, explain what is actually happening.
• It is not next-pixel prediction. The model starts from pure random noise and removes a little of it at each step, because it was trained to turn noisy images back into clean ones. Your prompt is embedded into a vector, the same kind you met earlier, and that vector steers every denoising step toward an image that matches, so a picture that never existed is what remains.
• One image is one denoise. Next, run the same trick over time: video is denoising over frames, and worlds you can act in.
• You can explain how image generators work: a denoiser walks from noise to an image while the prompt guides each step, and it is a different mechanism from next-token prediction.

Key takeaway

You watched noise turn into a picture step by step, and steered it: image generation is trained denoising, guided by your prompt's vector.

What you can do after this lesson

You can explain how image generators work, a denoiser walks from noise to an image while the prompt guides each step, and that it's a different mechanism from next-token prediction.

Prefer reading first? This lesson has a companion explainer: Diffusion models

Where it leads: One image is one denoise. Now run the same trick over time: video, and worlds you can act in.

This is the written summary. The lesson itself is interactive: you predict, drag and operate the mechanism above, and the reveal answers you.