// How AI generates images — from random noise to a masterpiece
💡 The core idea: Start with a completely random TV-static image. Then, in ~20-50 small steps, remove the noise guided by your text prompt. Each step asks: "given this noise and the prompt 'cat on a beach', what small change would make this look slightly more like a cat on a beach?" After many steps, the static becomes the image. This is diffusion.
Watch noise become an image
Prompt: "a golden cat sitting on a beach at sunset"
Ready — click Generate
0 / 6 steps
Denoising Diffusion
A diffusion model is trained by: (1) taking a real image, (2) adding random noise step by step until it's pure static, (3) training a neural net to predict and remove the noise at each step. After training, you can run the process backwards: start with noise and guide removal using a text prompt. The model learned how to "sculpt" images from noise.
20-50 DENOISING STEPSGUIDED BY TEXT PROMPTSTABLE DIFFUSION / DALL-E / MIDJOURNEY