Difference between huggingface demo and published pre-trained model

[Aki-Tomoya]

When I use the demo demo on huggingface for image segmentation, I get great results：

But when I use the published pre-trained model for segmentation, I get poor results：

Do you know what the problem is? What can I do to reproduce the results of the demo?