Tinker Diffusion models

by Samy (Kondasamy Jayaraman) · drawn with Colorfigures by April Zhu

The hinge drawing, clean.

Chapter 1 of 7

Noise to picture

Larkspur wants a picture of every part for its catalogue, drawn by a model instead of a photo shoot for each.

This chapter: How does a picture come out of pure noise?

About this chapter

You will learn

  • How forward noising buries a drawing in noise
  • How a denoiser runs it backwards, every pixel at once
  • How that differs from writing text a token at a time

Words you’ll meet

Diffusion model
A model that makes a picture by starting from noise and removing a little of it each step.
Noise
Random speckle added to every pixel. Enough of it hides the picture completely.
Denoiser
The part that guesses the clean picture behind a noisy one. This toy averages Larkspur’s twelve drawings, weighted by how well each fits.

Step 1 of 4

The darkroom: a clean hinge drawing lies in the tray. A makes pictures, but not the way a transformer writes text.

Predict first.

A transformer writes a reply one token at a time. How does a diffusion model make a picture?