Weird secret of AI image makers: before they can build a picture, they learn to destroy one. You add a tiny bit of TV-static noise, then a bit more, then more… until the picture melts into pure random fuzz. Grab the slider and dissolve a picture with your own hand.
Pick a picture, then drag the noise slider from 0% to 100%. At every step we mix in a little more random static. Same picture, slowly turning to fuzz — live, in your browser, one pixel at a time.
Every notch replaces a bit more of the real picture with gaussian static — the same kind of random fuzz used in real AI image models. Slide slow and watch the picture fight to survive. 🌫️
Real models don't dump all the noise at once — they add a little at each timestep, following a plan called the noise schedule. Press play and watch 12 steps go by, from crisp picture to total static. Numbers show exactly how much noise is in the picture at each step.
Here are three totally different pictures, all cranked to 100% noise. Press the button to re-roll the static. Can you tell which was the sunset? The smiley? You can't — and that's the whole point.
We noised this picture by a secret amount. Study it, then drag your guess slider to how much static you think is mixed in. Get within 15% of the truth and you score. Nail it 3 times to win the badge!
Here's the twist that trips people up: the forward process — the melting you just did — involves zero learning. It's fixed math. Add gaussian noise, follow the schedule, done. A calculator could do it.
Picture → static. Fixed recipe, no brain needed. This is just so the model has millions of "noise → clean" examples to study.
Static → picture. THIS is where a huge neural net trains for weeks to guess "what noise do I remove next?" The actual magic.
Same reverse trip, but a text prompt whispers "make it a dragon" at every step. That's text-to-image.
You learned forward diffusion: how models wreck pictures on purpose, one gaussian-noise step at a time, until every image becomes the same random static. That "everything melts to the same nothing" fact is the launchpad for making brand-new art.