Denoising Diffusion Probabilistic Models have a reputation for being math-heavy, and the papers don’t help — pages of variational bounds before you get to the one idea that actually matters: learn to reverse a known process, one small step at a time.
The forward process is the easy half
Take a clean image and add a little Gaussian noise. Repeat a few hundred times. Eventually you’re left with pure noise, and — critically — you know exactly how much noise you added at every step, because you added it yourself.
The reverse process is the whole trick
Train a network to predict the noise you added at each step, then subtract it out. Do that a few hundred times in reverse, starting from random noise, and you get a plausible sample from the training distribution.
For my thesis, the “image” wasn’t a photo — it was a radio path-loss map, and the network was conditioned on satellite imagery of the terrain. Same reverse-process idea, different modality. That’s the part that made it click for me: DDPMs aren’t really about images, they’re about learning to undo a corruption process you fully control during training.
The unglamorous truth is that most of the engineering effort goes into the conditioning signal and the noise schedule, not the denoising network itself.