Any particular reason why training with LPIPS loss?

[jotix16]

Thank you for the nice work :)

I was unable to extract the reason why the LPIPS loss is used during training.

It clearly makes sense with regard to the evaluation pipeline (i.e. directly optimize what you are evaluating for).

Could you tell whether LPIPS is necessary? Is it motivated empirically or just used?

Why not use the default loss ( l1-loss + l_ssim) that is used in the 3DGS seminal work?

[donydchen]

Hi, @jotix16. Thanks for your appreciation.

It seems like a general approach to use LPIPS loss in the feed-forward setting. I remember the visual quality would be a bit worse, and the quantitative scores would drop slightly, if the LPIPS loss was removed. You can verify this by re-training this project and setting the LPIPS weight to 0 at

mvsplat/config/experiment/re10k.yaml

Line 35 in ef976a1

weight: 0.05