Hi all, I've started just started using faceswap and am finding it somewhat frustrating to get a good end result. I keep ending up with blurry swaps.
I've read through some suggestions here and have done the following:
Made sure my source face to be placed is equal to or higher quality than the swap recipient
Ran model for 100,000+ iterations
Tried to get as many angles and lighting conditions as I could
It seems no mater what I end up with a blurry face so I have a couple questions...
I read somewhere that switching to no-warp is critical in the late stage training. My model is currently continuing now with "no-warp". Does no-warp do anything for blur related issues?
I am using the original trainer. However since I only care about A > B should I be using the unbalanced trainer? If so, where do I get the unbalanced plugin for faceswap, I seem to only have settings for Original in settings.
Is there something obvisous that I might not be doing as a newbie.
No-warp does make the image more clear if you enable it at the end of training, that's fine.
However, your biggest "issue" is likely that you're using the original model. Original is a 64x model, meaning that the images it produces are 64x64 pixels in size. This is quite small, especially when looking at HD sources. You should consider DFaker or Phaze A if your system can handle it. Know that training larger models is slower, but in exchange you get higher resolution and better looking results.