Denoising Diffusion
The exact recipe behind diffusion models: add noise on purpose, then train a network to take it back out.
Think of it like
Like deliberately smudging a drawing and teaching an apprentice to un-smudge it one pass at a time.
Example
The 2020 DDPM paper turned this into the standard training loop that powers today’s image and video generators.
How it actually works
You corrupt training images with Gaussian noise across many steps until they’re unrecognizable, then train the model to predict what noise was added at each step. Generation runs the same steps in reverse. The whole thing is a clever way to break an impossibly hard “make an image” problem into thousands of tiny, learnable “remove a little noise” problems.
For product teams
The underlying method you’re licensing whenever you buy into an image or video model.
For engineers
A latent-variable model where the forward process is fixed Gaussian noising and the reverse process is learned.
Related
- Diffusion Model — The model class this recipe builds.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome