DDPM loss derivation in a nutshell: ignore this, ignore that, predict the noise

ariG23498 · x · 2026-08-28

A developer jokes that the DDPM (denoising diffusion probabilistic model) loss derivation boils down to pages of "this term doesn't depend on it, safely ignore," "this can be ignored too," "this is really complicated, ignored" — until all that's left is figuring out how much noise was added. Paired with 😭🥀, it nails the pain of everyone who has wrestled with diffusion model math.

Original post →

More from Fun

Fun channel →