What an image costs in tokens
The same image through four vision-language designs. Change the resolution and watch each one spend the context window differently.
This scene needs WebGL. The controls and token counts still work.
drag to orbit
What to look for
- Two bars never move. The resampler hands over 64 latents and early fusion spends 1,024 codes at any resolution. Extra pixels are squeezed out before the language model sees them.
- Tiling follows shape before size. The tile grid is chosen by aspect ratio, so a small 3:2 image still becomes six tiles and a thumbnail. The count moves in steps of 256 and stops at 1,792.
- Native resolution pays for every pixel, twice over. Double both sliders and that lane holds about four times the tokens. Turn on attention cost and its square grows about sixteen times.