How Image Generation Works: Diffusion in Plain Terms
Most modern image generators are built on diffusion. The idea is so counterintuitive it sounds like a joke: to learn to draw, the model is first taught to ruin pictures.
Training in reverse
You take an image and gradually add random noise — step by step, until nothing remains but even static. The model is trained to predict exactly which noise was added at each step.
Having learned that, it can do the reverse: take pure noise and successively remove the noise it predicts. After a few dozen steps an image emerges from the static.
What steps are
The step count is how many cleaning cycles the model runs. Few steps give a raw, blurred result. Many give more detail but take longer with diminishing returns: the difference between twenty and forty steps is visible, between forty and eighty usually not.
Where text control comes from
Alongside the images, a text encoder is trained: it turns a description into a set of numbers, and the model learns to associate those numbers with the content of pictures. During generation the description sets the direction in which an image emerges from the noise.
An important consequence follows: the model understands not words but statistical links between words and pictures in its training set. If a concept appeared rarely there, no amount of rephrasing will summon it.
Why the same request gives different results
The starting noise is new each time, set by a number — the seed. Fix the seed and all the settings and you can reproduce a result exactly, which is useful when you need to change one detail while keeping the rest.
Go to the site →Read next
Adult Content Generation: What Models Can Do and Where the Limit Is AI Artefacts: Hands, Teeth, Jewellery and Other Weak Spots How to Phrase a Request So the Model Gives You the Frame You Want Local Models or Cloud Services: Which to ChooseThis material is intended for adults over 18 and is informational in nature.