Videos de denoising

How AI Makes Images From Pure Noise (Diffusion, Explained)

You type a few words and — seconds later — a picture appears that has never existed anywhere in the world. No clip art, no copy-paste, no artist. So how does an AI actually paint something from nothing? The answer is genuinely strange: it doesn't start with a blank canvas. It starts with a screen full of pure random noise — TV static — and removes it. This is the clearest possible explanation of diffusion, the idea behind DALL·E, Midjourney, and Stable Diffusion. No math required. We start with the twist that trips everyone up: image models don't "draw." They begin from a field of random static and, step by step, strip the noise away until a picture that was hiding underneath comes into focus. Creation by removing randomness. Then we unpack how that's even possible: • Trained in reverse. During training the model takes millions of real photos and slowly adds noise to each one, watching it dissolve into static. Do that a billion times and it learns to predict the exact noise that was added at every step. • The whole trick. If you can predict the noise that was added, you can subtract it. So to create a brand-new image, the model just runs the process backwards — from static, back toward a picture. • The denoise loop. Look at the noisy image, predict the noise, subtract a little, repeat — 20 to 50 times — each pass a little sharper, until only the image is left. • What decides the picture? Pure noise could become anything — a face, a forest, a bowl of soup. So your prompt gets turned into numbers the model understands, and at every single denoising step it nudges the guess toward your words. "A sunset over the mountains" pulls the noise, bit by bit, toward exactly that. The mental model to walk away with: the AI is a sculptor, the block of marble is pure noise, and your prompt is the chisel — every step chips a little randomness away until your image is all that remains. Chapters: 0:00 The picture that never existed 0:15 What AI image tools actually do 0:30 The twist: it starts with static 0:52 Watch noise become an image 1:11 How it learned — by destroying images 1:34 Adding noise, step by step 1:51 The trick: predict, then subtract 2:07 Denoising in a loop 2:26 But what decides the picture? 2:45 Your prompt steers every step 3:07 The mental model: a sculptor 3:24 Recap 3:45 Subscribe Making sense of AI, one concept at a time. Subscribe → @watchsuperintelligence #AI #Diffusion #StableDiffusion #Midjourney #DALLE #AIart #GenerativeAI #TextToImage #AIexplained #MachineLearning #ArtificialIntelligence #HowAIWorks