순방향 노이징과 역방향 디노이징

A diffusion model works in two directions: it slowly adds noise to an image, then learns to walk that process back to a clean picture. 🌫️

The forward process is fixed and needs no learning. It just keeps mixing tiny bits of Gaussian noise into your image, step after step.

After enough steps the image becomes pure random noise, looking just like TV static with no trace of the original left.

A timestep t says how far along the noising you are. Small t means a clear image; large t means almost total noise.

You can jump straight to any timestep by blending the image with noise in one shot, no loop needed.

A schedule decides how much noise each step adds. Early steps barely change the image; later steps drown it quickly.

The reverse process is the part we train: start from static and remove a little noise at each step until an image appears.

The clever trick: the network does not paint a picture. It predicts the noise that was added, so we can subtract it away.

Denoising happens over many steps, each undoing a small amount. Patience here is what makes the final image so sharp.

Because the forward process is known, you always have a perfect noise target. That makes the training signal clean and stable.

To create something new, you simply feed in fresh random noise and let the reverse steps sculpt it into a brand-new sample. ✨

What does a diffusion network actually learn to predict?

Diffusion adds noise forward and learns to denoise backward. Train it to predict noise, then start from static to generate fresh images. 🎉