Faceswap Forum

I know the answer is "as perfect as possible", but I'm hoping for some more qualitative answers, to help decide if something will be good enough, or will be a complete mess if I don't spend the hours adjusting it.

To give a specific example that I constructed from stock photos for illustration purposes:

Usually, you get a nice extract that looks something like this:

But every now and then, every frame from a scene looks kinda like this:

Will that still create decent results, or will the trainer think the model has a wide face with a hairy ear growing out of her cheek?

I'm thinking it's a gradual shift, so that's why I'm asking... how bad does this kind of thing need to be before it starts to really matter, assuming I'm not trying to convince someone that it's real?

In a similar vein, sometimes everything looks perfect, EXCEPT the eyebrows are like at the top of their forehead... how much of an impact will that have overall?

EDIT TO ADD: Given that the face that gets converted isn't really shaped like the actual mask or landmark lines (looks more like a square), I'm hoping the eyes and mouth are the "high priority" landmarks, and the rest of the image that would land outside the target's face area just gets clipped, so the bad edges don't matter that much.

If you are using a landmarks based mask (components/extended), then it matters a lot more, as the mask will be wrong if the landmarks are wrong.

If you are not using a landmarks based mask, you can disregard the jaw entirely. Only the core area landmarks are used for alignment (i.e. all landmarks except for the jaw).

"off" landmarks can effect the alignment in unexpected ways. Generally I'll fix a face with "off" landmarks within the manual tool and watch the face preview to see how much it impact the alignment of the extracted face. If it's significant, I will fix all those which are off, if the change is minimal I will not worry about it too much.

Correctly aligned faces = better training and better results.

Aah ok. Then I should probably keep fixing them, since I'm basically defaulting to extended.

It's funny though, even with a well-defined chin, videos converted with extended tend to cut off the bottom just below the mouth, rather than including the full contour of the chin.

Replicon wrote: ↑Wed Jul 28, 2021 1:21 am

It's funny though, even with a well-defined chin, videos converted with extended tend to cut off the bottom just below the mouth, rather than including the full contour of the chin.

That will depend on the centering used for training, the coverage used, and how much erosion is applied to the mask at convert time.