Grayscale architecture: Deepfacelab-SAEHDBW

[Twenkid]

The "issue" was the lack of a grayscale architecture - at least I didn't find one with an "easy" research and I haven't seen grayscale deepfakes. It seemed that everybody was making only color deepfakes(?) which is weird - black-and-white consumes 3 times less data, while it is stylish for portraits and there is less distraction from colors. It could be applied for classic movies as well.

I was about to ask if anyone had done this modification already or for directions, where to edit, but instead I decided to took the challenge and resolved it myselff: https://github.com/Twenkid/DeepFaceLab-SAEHDBW

To exemplify the improved performance, the model for the first test film "Arnold Schwarzenegger: The Governor of Bulgaria" is trained just on a Geforce 750 Ti 2 GB on Windows 10 (it gives only 1.45 GB out of 2 for programs).

It is a DFL-SAEHDBW architecture, df-udt mf 192x192 128x48x32x16. It was trained with a batch size of 6 most of the time, in the GPU. It could go 7 or 8, if not in GPU.

The files on disk initially are about 345 MB, which seems to be about the maximum size of a model that can be trained on Windows with this GPU(?). Other tested big models were e.g. df-ud 192 128x48x48x16 which resulted in the same size on disk. I pretrained also df-ud 256x96x32x32 281 MB, batch 4.

See more info and details in the repository.