Lesson 47 of 48
Diffusion
Adding noise on purpose, then learning to remove it.
Destroying structure is easy and exactly describable; the model only has to learn the reverse of a process you defined. That asymmetry is the whole trick behind image generation.
Do this
Take a mixture of two well-separated Gaussians. Implement the forward noising process, watch the structure disappear, then learn the reverse and sample from it.
The question that unlocks the next lesson
Why is the forward (noising) process in diffusion not learned?
- AIt is too expensive to learn
- BIt is defined by you in closed form — only the reverse, denoising step needs learning
- CIt is learned, just before the reverse process
- DBecause noise cannot be modelled