Diffusion · Star demo

Fuzz & Un-Fuzz

Draw a picture and bury it in static. Then watch Robo clean fresh static back into a picture, one step at a time.

Draw something on the left (or pick a doodle). Then slide the fuzz knob and watch your picture disappear!

Your picturestep 0
Robo's guess—
clear picturepure static

Robo's picture book

The only pictures Robo has ever seen. Bars = Robo's guess. Tap to hide or show.

How it works

Start with a picture

To a computer a picture is a grid of little colored squares. Each square is just a number.

Add static, bit by bit

Each step sprinkles a little more fuzz on top. After enough steps, it's only static. Messing up is easy!

Guess what's underneath

Robo looks at the fuzzy mess and guesses the clean picture. At first the guess is blurry. That's OK!

Clean a tiny bit, repeat

Robo wipes off a little static toward its guess, then guesses again. Step by step, a picture appears.

What just happened?

For grown-ups: what's really going on
  • Forward process. Each picture is a 32×32×3 vector x₀ in [−1, 1]. Noising follows the DDPM closed form xₜ = √ᾱₜ·x₀ + √(1−ᾱₜ)·ε, ε ~ 𝒩(0, I), with the cosine schedule ᾱₜ = cos²((t/T + s)/(1+s) · π/2), T = 100. The ε is fixed, so scrubbing the slider is smooth and deterministic.
  • The denoiser is exact, not a neural net. Real diffusion models train a network to predict ε (or x₀). Here the "training set" is the picture book, so we can write down the optimal denoiser: E[x₀ | xₜ] = Σᵢ wᵢ·xᵢ with wᵢ = softmax(−‖xₜ − √ᾱₜ·xᵢ‖² / 2(1−ᾱₜ)). The bars under each page are those wᵢ.
  • Sampling. "Smooth" is DDIM (η = 0, a probability-flow ODE discretisation); "Jiggly" is ancestral DDPM-style sampling (η = 1), which injects fresh noise each step. Very few steps make the early, blurry average guesses stick.
  • Why the guess locks on so fast. Pictures live in 3,072 dimensions, so even at fairly high noise the likelihoods of different book pages differ by huge factors and the softmax becomes nearly one-hot. Learned denoisers stay blurrier for longer.
  • Memorisation. Because the optimal denoiser for a finite dataset just reproduces training images, this sampler can only output pages from the book. Real models are deliberately imperfect (finite capacity, smooth inductive biases), which is what lets them generalise and make new pictures. Text-to-image models also add a text condition and work in a compressed latent space.
  • Further reading: Diffusion Explainer (Polo Club), Ho et al. 2020 "Denoising Diffusion Probabilistic Models", Song et al. 2021 "DDIM", Nichol & Dhariwal 2021 (cosine schedule).