I started using faceswap about a month ago. I tried ever possible setting (according to forum discussion) in phaze - A to get the best quality output with maximum setting enabled in 8gb VRAM (3070ti).
Before enabling central storage I tried subpixel and upscale hybrid for Decoder and hiddenlayer - FAIL
efficientnet V2 All @ 50 % - FAIL
I had to use "dny256" with (little modification) to get my training start @ batchsize of 8.
So basically:
I think with my current setup (8gb VRAM 16 GB RAM)and central storage enable. I can use subpixel, efficientnet v2 b3 @ 100% etc @ batchsize of 4 is a big deal.
After just 14 hours of training and just 23000 iteration the model looks phenomenal.
Details are awesome to say the least.
"I have attached my phaze - a, loss and global setting in this draft"
My only concern is when ever I am enabling GBLOCK my model get corrupted i.e I start getting solid colors from the very start itself.
Additionally if any one can suggest more improvement in my setting, suggestions are more then welcome.
CENTRAL STROAGE IS AWESOM.
Since direct storage is now a thing can we see faceswap utilizing this feature?