Articles
How does AI make an image? From noise to picture, plainly
An image generator does not assemble pieces found on the internet. It starts from random noise and cleans it into a picture, step by step. How it works and what you may do with the result.
A common belief is that an image generator searches the internet for suitable pictures and stitches them together. That is not what happens.
From noise to picture#
Most of today's image generators are diffusion models. During training they are shown pictures that have been progressively corrupted with noise until only random noise remains. The model learns to run that process backwards: given noise, produce slightly less noise.
When you ask for an image, the model starts from pure random noise and cleans it over dozens of steps — each step nudged a little further toward your description. What is left is a picture that never existed before.
Why hands used to come out wrong#
The classic failure, now largely solved, but it explains the mechanism well.
The model does not know a hand has five fingers. It knows what a hand statistically looks like. Hands appear in the training data in hundreds of poses, partly hidden, from every angle. A statistically plausible hand is not the same thing as an anatomically correct one.
The same logic explains garbled text in images, wrong reflections, and objects that are not physically possible.
What to put in the prompt#
Image generators reward specificity:
- Subject: what is in the picture
- Style: photorealistic, watercolour, technical drawing
- Composition: close-up, from above, wide shot
- Light: soft morning light, hard shadow, backlit
- What to avoid: many tools support a negative prompt
"Coffee" gives you a generic picture. "Close-up of black coffee in a white cup on a wooden table, soft side light from a window, shallow depth of field" gives you the picture you may actually have wanted.
Copyright: three things to know#
The training data is contested. Models were trained largely on images scraped from the internet, often without permission. Litigation is ongoing in several jurisdictions and there is no settled answer yet.
Ownership of the output is unclear. In many countries a purely machine-generated work gets no copyright protection, because there is no human author. Which means your generated logo may not be protectable.
Commercial use depends on the service. Read the terms. Some permit commercial use, some only on paid tiers, some not at all.
A practical rule: a generated image is a good illustration and a bad trademark.
Where it works best#
Illustrations, drafts, mood boards, visualising an idea. Anything where the picture has to be good enough rather than exactly right.
Where it works badly: anything where the detail has to be true. Product photography, technical drawings, anything that states a fact.