What the Model Sees is a guided demo. Pick one of eight images from this site, press Scan, and a vision-language model shows what it made of it, task by task: captions, objects, labelled regions, even text it thinks it can read.
The results are recorded. I ran Florence-2 on every image ahead of time, so there is nothing to download. It turns the image into tokens, the units a language model reads, and writes its answer one word at a time.
A last step turns those notes into an image prompt, written without ever seeing the photo.
It never says how sure it is, and it can be wrong.