Tinker256 × 256 px: 65,536 positions × 50 steps = 3,276,800 units of work.
Chapter 6 of 7
Why it costs what it costs
Larkspur’s catalogue has thousands of parts, and finance wants to know what a picture costs before anyone presses go.
This chapter: What makes a diffusion picture expensive, and how do newer models cut it?
About this chapter
You will learn
- Why the work grows with steps and with picture size
- How latent diffusion shrinks what the denoiser reads
- How few-step models cut the steps instead
Words you’ll meet
- Position
- One place in the picture the denoiser reads each step: a pixel, or a cell of the latent.
- Latent diffusion
- Running the denoiser on a small encoded copy of the picture, then decoding the result to full size once.
- Few-step model
- A model trained to make a good picture in one or a few steps, often distilled from a diffusion model.
Step 1 of 4
Each step, the denoiser reads every pixel of the picture. Larkspur’s toy prints are tiny; catalogue images are not.
Predict first.