Training loss difference

[Germany321]

Hi, I trained the model with VGGT encoder based on the initial version of your code(before refactoring model) and noticed that the training loss is larger than the model without VGGT. Is this consistent with your empirical record?

[zd11024]

Yes, your observation aligns with our expectations. For 3D tasks, the training loss with VGGT is smaller than that without VGGT, indicating that incorporating VGGT enables the model to fit the training data better.

[0x4e8]

he: loss(VGGT) > loss(no-VGGT)
you: loss(VGGT) < loss(no-VGGT)
😳