Image generators such as Stable Diffusion are built on diffusion models: they learn to turn random noise into structure, one small step at a time. Drag the slider or press Denoise to watch it happen.
How it works
During training, real data x₀ is mixed with Gaussian noise ε: xt = √ᾱt · x₀ + √(1 − ᾱt) · ε, where ᾱt goes from 1 (clean) to 0 (pure noise) over the steps t = 0…T. A neural network learns to predict the noise contained in xt. To generate something new, the model starts from pure noise and removes a little of the predicted noise at every step.
What is simplified here
This demo already knows the target shape and simply moves along the same formula, with a cosine noise schedule and 50 steps. A real model has to estimate the noise at each step from what it learned during training: that is why it produces new samples instead of copies. In astrophysics, diffusion models are being explored for mock galaxy images, source deblending and the reconstruction of noisy data.