AI image generators are just really good at removing noise

When you ask an AI to "paint a cat riding a bicycle," it doesn't sketch from scratch the way a person would. Most modern image generators — like the models behind popular art tools — use a technique called diffusion, and the core idea is surprisingly backwards: they learn to create by learning to destroy.

During training, the model takes millions of real images and repeatedly adds random noise — those TV-static speckles — until each picture becomes pure visual mush. Then it learns the reverse: given a noisy image, predict what the slightly-cleaner version underneath looked like. Do that prediction thousands of times and the model becomes an expert at denoising.

To generate something new, the model starts from a canvas of pure random noise and denoises it step by step, nudging the pixels toward a coherent picture. Your text prompt acts as a steering signal — a guidance that biases each denoising step toward "cat," "bicycle," and so on. The final image was never copied from anywhere; it condensed out of static, shaped by the patterns the model absorbed during training.