Add noise gradually
Diffusion models define a forward process that gradually corrupts data with noise. A model learns a related reverse process or prediction task.
Generate by reversing
Generation begins from a noise distribution and applies a sequence of learned denoising steps. Different diffusion formulations and samplers make different choices. The two contributions are added to obtain the observation; this is a linear-blend illustration.
INTERACTIVE EXPLORATION
Blend signal and noise
For observation (1 − t) × signal + t × noise, change t. Signal is 1 and fixed noise is −1.
One noisy number
The experiment blends a fixed signal and fixed noise. It illustrates how a noise amount changes an observation; it does not train or run a diffusion model.
Check your understanding
Does a single noisy example constitute a diffusion model?
No. A diffusion model requires a defined process and a learned reverse or prediction mechanism.