Lesson 3 • 3 min
The Clever Trick
Why removing noise is easier than creating from scratch
The dirty window insight
Imagine you have a dirty window. What's easier: cleaning the dirt off to reveal what's behind, or painting the view from scratch on a blank canvas?
Diffusion models use this exact insight. Instead of learning to generate images from nothing (incredibly hard), they learn to clean up noisy images (much easier).
Drag the slider to see how noise is progressively removed to reveal an image
But wait—where do we get dirty windows to practice on? We make them ourselves! During training, we take clean images and add noise step by step until they become pure static. Then we teach the model to reverse each step. The model learns: "Given this noisy image, what should the slightly cleaner version look like?"
The forward process adds noise, the model learns to reverse it
Quick Win
You now understand diffusion's core insight: it's easier to learn denoising (cleanup) than generation (creation from scratch).