Making images from text — and what can go wrong

Write «a fox reading a book in the forest, watercolour» — and a model creates the image. It does not copy an existing picture; it builds from noise, step by step, towards something that sits close to your text on the map of meaning.

Three things to watch out for:

• Mistakes: hands, text inside the image, counts — the model knows how things usually look, not the rules.
• Stereotypes: «a doctor» often gives a certain kind of person, because the training data looked like that.
• Copyright: the model was trained on millions of images — often without the creators being asked. Imitating a named artist's style is contested.

An experiment (with a free image generator):

1. «a doctor» — describe the person you got. Try five times. A pattern?
2. «a doctor, an older woman, a hospital in Nairobi» — what changed?
3. «a hand holding six apples» — count the apples and the fingers.
4. «a sign with the text STOP» — did the letters come out right?

Write it down: what does the prompt steer well, what often goes wrong, and what do you think about the model having learnt from other people's pictures?