UNet3D consistency with MegaPortraits

[MikeRogers27]

Hi neek2302

Thanks for publishing this repo. I'm trying to follow the code and understand how it corresponds to your paper and the MegaPortraits paper. I'm using this MegaProtraits paper here as a reference to the architecture: https://arxiv.org/pdf/2207.07621

In the implementation of UNet3D that I think corrsponds to G3D in the paper, I noticed that the input does not flow all the way through the UNet architecture in the way indicated in the Appendix here.

In the code there is a learnable parameter that replaces the input at the lowest resolution level.

Is this something that came from the original MegaPortraits authors, because it seems pretty different to the block diagram above?

Thanks!

[johndpope]

might help - https://github.com/johndpope/MegaPortrait-hack

[neeek2303]

Hey @MikeRogers27 Unet3D architecture-wise fully correspond to the one that was used in MegaPortraits. I don't remember fully details, but I believe just number of channels in some layers might be different