Most popular image generators use a technique called diffusion. It sounds complex, but the core idea is simple.
Learning from images and captions
The model is trained on huge numbers of images paired with text descriptions. It learns connections between words and visual features: "golden hour" means warm light and long shadows, "watercolor" means soft, bleeding colors.
From noise to image
During training, images are gradually covered with random noise, and the model learns to reverse the process, removing noise step by step. When you generate an image, it starts from pure noise and removes it bit by bit, guided by your prompt, until a picture emerges.
Why prompts matter
Your words steer every step of the denoising. Specific descriptions of subject, style, lighting and composition give the model clearer guidance.
Why hands and text used to fail
Small, precise structures with strict rules (five fingers, correctly spelled words) are hard to learn from patterns. Newer models are much better but still make mistakes.
Seeds and variation
The starting noise is set by a "seed". Different seeds give different images from the same prompt.