DDPM & DDIM

DDPM — Denoising Diffusion Probabilistic Models, 2020 — made the idea concrete and trainable. Training is almost comically simple: take an image, pick a random timestep, add that much noise, ask the U-Net to predict the noise, penalize the error. Repeat millions of times.

Sampling is the reverse: start from pure noise, apply the denoising step ~1,000 times with a sprinkle of fresh randomness at each. Slowly — remarkably — a photograph condenses out of static.