Worse results than on Meta's web demo

[holma91]

Hey, I'm somehow getting worse results with this in auto1111 than what I get on Meta's web demo at (https://segment-anything.com/demo). I'm using the h model which is the largest, which I assume is what they are using aswell. Has anyone else noticed this?

Here is one example, when I do one positive point for the shirt and one for the pants, along with a negative point on the face. First is from Meta's demo, and second is in auto1111:

This is just one example, it's worse in the same way for every image I try this on. How is this possible, shouldn't the same model be used in both instances?