Page 1 of 1

Any new training destabilizes faces

Posted: Wed Jul 26, 2023 8:38 pm
by StarkRaven

Hello,
Hope all is well. Having an odd issue. Mainly using villian. It started after an update about 6 months ago. Any model I train, new or existing becomes unstable. The face bounces or distorts, slightly changing size in the converted video.
If I use any model that was trained before this time the converted video is great. I just can't update the model for the video I am trying to convert or the face wobbles.
I have unistalled Faceswap and all related Cuda libraries multiple times but the problem persits. I have re-extracted my training faces. I have tried training both with and with out masks.Nothing corrects the problem. I usually train for 750,000 to 1,500,000 iterations. Even supplimental training of simply 10,000 iterations on one of my previously trained models cause the effect.
Any thoughts or ideas would be greatly appreciated. I have been using Faceswap for about 2 years and still get great results, as long I am using any model trained over 6 months ago.
I have a sample video I can upload to demonstrate both the stable older models and the newer wobbling one. Same video, same extraction, mask (bisenet), and alignments the only difference Is on what version of faceswap and cuda the model was trained.
Thanks,
and all the best.


Re: Any new training destabilizes faces

Posted: Thu Jul 27, 2023 11:13 pm
by torzdf

This sounds very odd.

A sample video would be appreciated.


Re: Any new training destabilizes faces

Posted: Sun Jul 30, 2023 1:09 am
by Bigbobbigybig

I am having the same issue. It was running fine a maybe a week ago. Now all of the faces are coming in blurry, distorted, or imperfect. I was going to try and uninstall is and retry the program but I'd love to know what happened.