Point cloud quality degrades when inputting camera intrinsics/extrinsics

I am using the Giant model. I tried the web demo and the results were excellent. However, when running it locally, I attempted to input the camera intrinsics and extrinsics expecting better results, but the outcome was actually worse . All other parameters remained consistent.
Here are the results:
Without camera parameters:

With camera parameters:

Is this behavior normal?
If I do not provide intrinsics and extrinsics, does the generated point cloud possess metric scale ?

Reproduce the Giant model locally with the same inputs, comparing runs with and without camera intrinsics and extrinsics. Check how the local demo accepts and applies those camera parameters, then determine whether the observed degradation and the metric scale of parameter-free output are expected.

Written by the indexing model from the issue text.