TinkerPASS1,600 px: 2,610 image tokens; 5 of 5 fields read right.
Chapter 2 of 7
Size and detail
An invoice at full size costs thousands of tokens. Larkspur gets hundreds of invoices a week, and someone suggests sending them smaller.
This chapter: What do you lose when you send a smaller image?
You will learn
- Why fewer pixels means fewer patches, and so fewer tokens
- Why small print is the first thing to go
- Why a very large image is scaled down before it is read
Words you’ll meet (3)
- Resolution
- How many pixels an image has across and down. Here the size is the long edge in pixels.
- Downscale
- Shrinking an image to fewer pixels. The model’s host does it to any image over its limits.
- Small print
- The smallest text on a page: terms, footnotes, fine detail. It blurs away first.
The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.
Step 1 of 3
At 1,600 px the invoice costs 2,610 image tokens, and the reading gets every field right. Fewer pixels means fewer patches, so fewer tokens. Before you shrink it: