Advice on settings for Hi Res Results

Want to understand the training process better? Got tips for which model to use and when? This is the place for you


Forum rules

Read the FAQs and search the forum before posting a new topic.

This forum is for discussing tips and understanding the process involved with Training a Faceswap model.

If you have found a bug are having issues with the Training process not working, then you should post in the Training Support forum.

Please mark any answers that fixed your problems so others can find the solutions.

Locked
User avatar
MattB
Posts: 22
Joined: Fri Aug 19, 2022 4:54 pm
Been thanked: 5 times

Advice on settings for Hi Res Results

Post by MattB »

I'd like advice on the best configuration for high resolution face swaps. I'll have roughly 15K images of each face extracted at 1024 from hi-res photos. I did prep some of the photos using GFP-GAN. I assume Phaze-A, but which training preset to start with and what settings to toy with. The end result is what matters, not the time it takes to run the models. I should have 40GB of VRAM to play with, so the model can be pretty large.

Does a smaller batch size and shallow learning curve sacrifice time for a better result? I've read that increasing filters in Phaze-A is the best use of VRAM, but not sure where to start. Also, since photos don't present a consistent "range" of faces should I use instance normalization rather than batch normalization?

So many settings in Phaze-A it makes my shot glass of a brain overflow.

User avatar
torzdf
Posts: 2636
Joined: Fri Jul 12, 2019 12:53 am
Answers: 155
Has thanked: 128 times
Been thanked: 614 times

Re: Advice on settings for Hi Res Results

Post by torzdf »

Unfortunately I don't have any definitive answers for you, however, there are some user supplied tests and presets in the Phaze-A thread which should help you get started.

MattB wrote: Wed Feb 08, 2023 9:28 pm

I did prep some of the photos using GFP-GAN

Generally feeding AI upscaled images into the model can lead to issues, so be aware.

MattB wrote: Wed Feb 08, 2023 9:28 pm

Does a smaller batch size and shallow learning curve sacrifice time for a better result

Generally, yes. You can start with a higher batch size to get the model going, then lower it though.

Last edited by torzdf on Thu Feb 09, 2023 12:02 pm, edited 1 time in total.

My word is final

Locked