What the Model Sees

Pick a photo, then press Scan. A vision-language model has been run over each photo on every task it knows. You will see what each one made of it.

How sure is it?

It does not say. The model writes the most likely words, one after another, and gives no score for any of them. A caption can sound certain and still be wrong, and boxes can land on the wrong thing or name something that is not there. Text it reports is often not in the picture at all. Check it against what you can see.

Recorded results: Florence-2 ran on these photos ahead of time, so there's nothing to download.