Diffusion Models: Learning to Undo Noise
Training adds noise to a sample in small steps and asks the network to predict the noise mixed into the result. Generation runs that chain backwards from pure noise, so one hard sampling problem becomes a long series of easy denoising ones.
What a learner can do afterwards
- Describe the forward corruption process and say why it needs no training
- State what the network predicts at each step and how repeating it yields a sample
- Say what more sampling steps buy and what they cost
1 · Read
A diffusion model is a generative model whose hidden code is a noisy version of the sample itself. Training starts by wrecking a sample on purpose. Noise is added in small steps until nothing but static remains. This forward corruption is a fixed recipe, so it needs no training at all.
Every forward step mixes in fresh noise, so the noisy sample at any level is the clean sample blended with one lump of noise. At each level the network is shown the noisy sample and predicts that whole lump: the noise mixed in so far, not just the last small dose. Generation runs the chain backwards: start from pure noise, remove a small part of the predicted noise, and repeat until a clean sample stands, with the usual recipe stirring a little fresh noise back in at every step but the last. One hard sampling problem becomes a long series of easy denoising ones.
More sampling steps usually buy a cleaner sample and always cost more waiting. Fewer steps run fast but can leave rough edges and odd textures.
When results look noisy, add steps before touching anything else. When results look fine but slow, trim steps first.
Learn each denoising step, then walk backwards from noise to sample.
2 · Watch
Take it off screen
Where it sits
This opens up
Nothing builds on it yet.
8 questions wait behind this lesson, each with its answer explained. Every answer feeds the sky: stars light as they are learned, and dim when it is time to come back.