search

ai-forever / ghost

A new one shot face swap approach for image and video domains
Apache License 2.0
1.25k stars 265 forks source link

any way to improve faceswap video during close up? #28

Open Funkybo0dah opened 2 years ago

Funkybo0dah commented 2 years ago

I find the model is more effective in faceswap for video during medium or long-shots, but it really breaks down during close up shots (where the person's face takes up most of the screen).

Any advice on how to improve this?

I find that even when I increase inference rate to 60, it does not improve much.

Thanks!

JohnnyRacer commented 2 years ago

I believe this is an issue with the model not being able to align the face properly on close up shots of faces due to a lack of detail for the boundaries of the face. I think there is a way is get the model to run properly if you fine tune on a dataset containing just close ups so the model can learn more accurately on how to extrapolate the positioning without having clear face boundaries in view.