Use with existing ply files?

[antithing]

Hi, can the pbr and relighting approach work with existing Gaussian ply files?

[JinchenBai]

I think it can't, two reasons:

1. The R3DG need normal attributes which is absent in the vanilla gaussian, and need to be trained use the L_n mentioned by the paper
2. The distribution of gaussian is different which maybe cause error normal estimation , while in this paper, the R3DS are contrained to the surface by the L_u

[HeliaCheng]

yes it can. 1. you need to render image from this ply as gt. 2. first run -t render then -t neilf bcs in the first round of training you will get normal for each gaussian ball. 3.and yes, the final results may not as good since the noraml est is not good enough. as my result shows, the material sometimes looks very wrong and get weired color