Ask most people how an AI image generator works and they will describe something like a very fast artist: you type a prompt, the machine imagines a picture, and it draws it. That is a reasonable guess, and it is wrong in an interesting way. The technology behind tools like Midjourney, Stable Diffusion, and the image and video features baked into the big chat apps is the diffusion model — and it does not paint anything. It removes noise.

The counterintuitive core

A diffusion model builds an image by starting from a field of pure random static — the grey snow an old television showed with no signal — and gradually cleaning it up. Over a series of steps it takes away a little noise each time, and a picture slowly emerges from the fog. It sounds like alchemy. The trick is in how the model learned to do it.

How it is trained: learning to undo the mess

Training runs the process backwards. You take a real image from the training set and add a small amount of random noise. Then a bit more. Then more again, over many steps, until the original is completely buried and nothing but static remains. That is the easy direction — anyone can add noise. The model's job is to learn the hard direction: shown a noisy image, predict the noise that was added so it can be subtracted. Do that across millions of images and billions of noising steps, and the model builds up a statistical sense of what real images look like at every level of corruption.

How it generates: the same skill, in reverse

Once trained, generating is just that skill pointed at random static. Start with pure noise — which the model treats as a maximally corrupted image — and ask it, step by step, to remove the noise it thinks is there. After twenty to fifty passes, a coherent image is left behind. Because the starting static is random, the same prompt produces a different picture every time.

Where the prompt comes in

The text prompt steers the denoising. Your words are turned into numbers by a separate text encoder, and those numbers are fed into the denoising network at every step, nudging each cleanup toward the description. Ask for "a red bicycle at sunset" and at each step the model removes noise in the direction that makes a red bicycle at sunset more likely. The image is not retrieved from a library; it is pulled out of the noise by the prompt.

Why it can run on a laptop: latent diffusion

Doing all of this at full pixel resolution is slow and expensive. The breakthrough that put image generation on consumer hardware — the approach behind Stable Diffusion — was to run the whole noising-and-denoising process in a compressed "latent" space: a smaller mathematical representation of the image, decoded back to full pixels only at the very end. Same idea, a fraction of the compute.

The trade-off, and video

The price of the quality is that diffusion is iterative. Each image is many passes through the network, which is why a generation takes seconds rather than being instant, and why so much research goes into cutting the number of steps down. Video is the same idea with a harder constraint: the model must denoise not a single frame but a sequence that stays consistent from one moment to the next. That extra demand is a large part of why AI video is so much more compute-hungry than a still image.

What this tells you about the technology

Two things are worth holding onto. First, a diffusion model does not work by storing its training images and collaging them together — it stores a learned sense of what images look like, though researchers have documented rare cases of a model reproducing a specific training image almost exactly. Either way it is unavoidably shaped by what it was trained on, which is exactly why questions of consent and provenance in training data are real and still unresolved. Second, there is no imagination in the loop and no understanding of what a bicycle is. There is a very good statistical guess about which pixels of noise to take away. That the process produces something striking, or occasionally something with six fingers, both follow from the same fact.