How do u guarantee the prediction mesh and gt mesh in the same scale?

[ForeverRuri]

As i know we first predict voxel in some pre-define size like 48*48*48 and then with the help of cubify to get the mesh without refinement.
But when we cal the loss like chamfer or normal loss, we need the prediction, gt pair in the same scale otherwise the two losses depend on coordinates are meaningless.
I know change the value of focal length and t_c(center in z axis) will change the cords, but i not find them in training process.
Can u tell how what's wrong?