Introduction
The previous topic explained the big idea: noise images in a fixed way, and learn to undo it. This topic makes it concrete:
- training turns out to be remarkably simple: add noise to an image, ask the network to guess the noise, and use mean squared error;
- sampling starts from pure noise and repeatedly applies the network, removing a little noise each time;
- faster samplers reduce the hundreds of steps to a few dozen or even fewer.
Think of training as a series of "spot the difference" puzzles: we show the network a picture with static added and ask it exactly which speckles we added. Sampling is then doing that hundreds of times in a row, starting from a picture that's nothing but static.