Why it matters

DDPM is diffusion foundation. Understanding shapes theory.

Advertisement

The architecture

Forward: q(x_t | x_{t-1}) = N(sqrt(1 - beta_t) x_{t-1}, beta_t I).

Reverse: learned p(x_{t-1} | x_t).

DDPM mathForward Markovadd noiseReverse MarkovdenoiseLosssimplified MSEHo et al. 2020 established diffusion for generative modeling
DDPM.
Advertisement

How it works end to end

Ho et al. 2020.

Simplified loss: MSE on predicted noise.

Beta schedule: linear or cosine.

1000 steps foundational.