Tinker Diffusion models

by Kondasamy Jayaraman · drawn with Colorfigures by April Zhu

Drawing the pictures…

Chapter 5 of 7

Steer with a label

The catalogue needs a drill on the drill page. Leaving the part to chance won’t do.

This chapter: How does a label steer what the model draws?

About this chapter

You will learn

  • How a label changes the denoiser’s guess
  • What the guidance scale does, from 0 to 4
  • Why stronger guidance costs twice the predictions

Words you’ll meet

Label
The part asked for, given to the denoiser with the noisy picture. Also called a class or a condition.
Guidance scale
How hard to push toward the label: 0 ignores it, 1 follows it as trained, more pushes past it.
Classifier-free guidance
Mixing the denoiser’s guess with the label and without it, then pushing further in the label’s direction.

Step 1 of 3

8 pictures from 8 seeds. The card asks for a drill; the says how hard to push.

Predict first.

At guidance scale 0, how many of the 8 pictures come out drills?