Stable Diffusion: how does an AI pull a picture out of pure noise?

An image AI like Stable Diffusion doesn't draw. It starts from pure noise, and wipes it away… until a picture appears. How? In 2015, researchers published a method that learns to remove noise step by step, inspired by statistical physics. In 2020, researchers at UC Berkeley showed that this diffusion approach can produce good-quality images starting from pure noise. In August 2022, Stability AI released Stable Diffusion, with its code and model weights shared under an open licence. Three steps: add noise to learn, remove noise to create, and guide with text. Step one: learn by adding noise. The model sees huge numbers of captioned images, each gradually covered in noise. At every dose, it must guess the noise that was added. Bit by bit, it learns what an image looks like underneath. Step two: create by denoising. Start from pure random noise; the model predicts the noise, removes a little, and repeats. Step after step, shapes appear, then details. Stable Diffusion 3 even learns a straight path from noise to image: rectified flow. Step three: text guides. The prompt, say “an artist in her studio, evening light”, becomes numbers. At every step, those numbers guide the cleanup. One setting decides how closely to follow the prompt: too little, it drifts; too much, it overcooks. Stability AI has rivals: OpenAI's DALL·E, Midjourney and Adobe's Firefly. At KOAEE, our assistants do the opposite: they generate no images. In a restaurant, a hotel or a shop, they grasp what the question means, and show the business's own catalogue photos. Our clients' customers see the catalogue, never a picture dreamed up by the machine. To understand AI in 150 seconds, subscribe to the channel. The AI that stays between you. Try it at koaee.ai.