Diffusion on Cats, Faces, Baseball Cards

The idea of degradation is harder to conceptualize than the GAN, I think. The usefulness of doing it this way is that you could start with noise and build something from that noise. If I train diffusion on cats, faces, baseball cards, and I ask for a cat face baseball card, it can connect all these things. So in essence, what's interesting about diffusion that you don't necessarily get with GANs is that steerability.