Focal lengths in estimated intrinsic are in similar ratio as input images' aspect ratio

Hi, first of all, impressive work! Didn't expect to see anything claiming to surpass vggt in such short period of time and there you are. Congratulations.

I'm puzzled by the predicted focal length though. For input non-square images, the predicted intrinsics typically have fx and fy to follow roughly the aspect ratio of the input image, as if they were derived from a NDC projection matrix instead of a pin hole camera projection. I wonder though whether the input argument expects similar NDC projection matrix too, if I have ground truth camera parameters?

Thanks