Open kalle07 opened 1 month ago
Yes, this the issue The best result - when faces are in the "proper" position (have no tilt > 90)
The only solution (unless it is a CloseUP) is to flip your image so it is in or close to Front Portrait position. Swapping works well even at about 90 degrees from the ideal position. For upside down, just rotate the image. Photos of Faces rotated in close to or away from camera in Z dimension, become problematic as the angle grows and there is no solution.
but what is the problem ... inswapper_128.onnx ? or is it the implementation to search for pattern in all rotations ?
I think you've hit the nail on the head: inswapper and the like. Perhaps you have noticed that we humans also have some problems with upside down faces. It takes us longer to recognize the face. It is even more of a problem with negatives (light and dark is reversed). In fact we are terrible with negatives. This does not mean that Ai should have a problem, in our applications, because it can be solved, as you already suggested by preprocessing rotations (flipping in case of upside down faces, and so on.) Evolution has not had to make us work with negative images, but Ai could fix that easily by just reversing the shading.
First, confirm
What happened?
dont work in upside-down positions faces much to big and distorted
all works fine wiht faces more or less in right position
Steps to reproduce the problem
-
Sysinfo
win10 latest version a1111 and latest reactor
Relevant console log
Additional information
No response