Loading slide
Loading contents...
A language model builds its output the only way it can: one piece at a time, left to right, each piece final the moment it lands.
Image generators do not work this way at all. They do something that sounds, at first hearing, absurd.
They start with static. Pure random noise, the visual equivalent of an untuned television, containing nothing.
Then they remove a little of it. Not all at once. The model looks at the noisy field and predicts what a slightly less noisy version would look like, given the description you asked for. It produces that slightly cleaner version. Then it does it again. And again, perhaps twenty or fifty times.
Shapes begin to loom out of the static. Then edges firm up. Then details. After enough passes, an image.
The training is what makes this less strange than it sounds, and it is a lovely trick. Take a real photograph. Add a little noise. Now train a network to undo exactly that, to take the noisy version and recover the original. Do this millions of times, at every level of noise from barely speckled to complete static.
What you end up with is a machine that is very good at one narrow task: given a noisy image, make it slightly less noisy. Run that machine on pure static and it will hallucinate structure into the emptiness, because removing noise is all it knows how to do. Steer it with a text description, and the structure it invents matches your words.
The technical name is diffusion, borrowed from the physical process of something spreading out and dissipating. Training learns the reverse route.
The chain took years. Researchers described diffusion models for generation in 2015. A 2020 method made the results much stronger. Latent diffusion then performed the expensive denoising in a smaller learned representation, which helped make high-resolution text-guided image generation practical.
An image is not written. It is developed, the way a photograph emerges in a darkroom tray, out of a field of noise that never contained it.