Faceswap Forum

d3x

So you need 200,000 iterations to do a half decent job then for a 40 second video? Wow. 200K iterations is a totally arbitrary number, it can take a very short or very long time to get there depending on many different factors such as model, learning rate, loss functions & batch size It's just ...

d3x

I thought I would get better results from VG16, but it feels like each got its pros & cons With VG16 the moire pattern is a lot less noticeable when training at high values (like 50%) or sometimes even gets completely trained out with lower values (even when zooming in to 200%) With Alex I get a...

d3x

for some reason it was just running far behind on training for 1 side of the face in certain angles, like 1 side was perfect and after a long long time it started matching the other side with the same proportions probably due to a lack of B training data for those angles, and it took a long time to ...

d3x

yea it turned out it was just nowhere near done on training the teeth, i got much better eventually

d3x

Apparently it was LPIPS making it impossible to resolve the teeth on the swapped face even after 2mill iterations (Phaze-A dny512 preset) Once I removed it from the loss functions, the model finally got the teeth right after another 100k iterations - but I did lose all the sharpness I gained from LP...

d3x

Teeth are always hard to get right for me... is there any trick to this? or certain trainers that perform better for teeth? It seemed to be less of an issue when I was working with lower output size models, but since I started doing 256px-512px output it just takes forever to get the teeth right whi...

d3x

Has anyone experienced this? I got a dataset which, when trained long enough, always ends up with a difference in size and color between left & right eye on some of the converted faces (right eye is slightly larger and sometimes a more blueish tint or sometimes a bit cloudy compared to the left ...

d3x

I find that leaving blurry faces in helps with initial training, but I try to remove most of them for the final "no warp" training to bring out more detail (though I'm not sure if this actually helps much)

d3x

135k iterations is nothing 😅 Seems very undertrained With villain (or any trainer with high resolution output) it always takes me a couple million iterations before it gets all the details right Sidenote: when converting you can increase the Face Scale in the Face Processing section to better line u...

Faceswap Forum

Search found 9 matches

Re: [Tip] Why you need more sources (with examples)

LPIPS_ALEX vs LPIPS_VG16

Re: difference between left & right eye

Re: Teeth

Re: Teeth

Teeth

difference between left & right eye

Re: Should I remove blurry faces from extracts?

Re: Newbie - should I continue with this data/training