🖼️ How does AI generate images?

Image AI works by learning to destroy and rebuild images. First, the model is trained to add noise to a picture until it's pure static, then reverse the process — to take static and turn it back into the picture. It learns what makes a 'cat' a cat by repeatedly rebuilding cats from noise.

Once it can turn noise back into images, you can steer it. Give it a text prompt ('a cat astronaut'), and it starts from static and rebuilds — but this time the rebuild is pulled toward whatever the prompt describes. The text guides the noise-removal toward a cat astronaut instead of just any cat.

That's why image AI is often called 'diffusion': the image 'diffuses' into noise and back, and your prompt nudges the direction. Each generation starts from random static, which is why you get a different image every time — and why the same prompt never gives the same result twice.

The result is genuinely new images that never existed, built from patterns the model learned. It's remarkable — and it's why AI images can also get details wrong (hands, text, fingers), because it's predicting what *looks* right, not what *is* right.