The core trick

Modern image generators — Midjourney, DALL·E, Stable Diffusion, FLUX — are almost all diffusion models, and they work in a way that sounds backwards at first: they learn to remove noise. During training, the model is shown real images with progressively more random static added, and it practices reversing each step — predicting what the slightly cleaner version looked like. Do that millions of times and it becomes an expert at turning noise into structure.

From noise to picture

To generate a brand-new image, the model starts from a canvas of pure random noise and denoises it step by step — usually 20 to 50 small steps — each one nudging the static a little closer to a coherent picture. There's no database of images being copied; the picture emerges from the noise, shaped entirely by what the model learned about how images look.

Why your prompt steers it

Left alone, the model would denoise toward some plausible image. Your text prompt is the steering wheel. The prompt is converted into an embedding — a numeric representation of its meaning — that biases every denoising step toward pixels matching your words. "A red bicycle" tilts each step toward red, round, spoked shapes.

This is why specific prompts work better: a vague prompt gives the model a vague target, so it fills the gaps with its own averages; a detailed prompt gives it a precise direction to pull toward at every step.

A diffusion model doesn't paint a picture and it doesn't fetch one — it sculpts away noise until an image appears, with your prompt deciding which way to carve.

Try this: Generate the same prompt twice with different starting noise (most tools call this the "seed"). You'll get two different images that both match your words — vivid proof that the picture is built fresh from noise, not retrieved from a library.