1,5 fps Using GFPGAN??? is possible? #135
-
|
Hi, I have a computer I9 i12900k, whit 64GB Ram and a Graphic Card RTX3060 with 8GB Vram. Thanks. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
tldr; The current enhancers except DMDNet don't allow parallel execution. Yes it's possible and there's nothing wrong. The problem is that all enhancers in their default form don't like running in parallel. This means that your face swapping is properly threaded but after that every thread has to wait in order for post-processing. Also every face restoration does the face recoginition process again. I know this is stupid but to improve this you would have to dissect the existing code and rewrite parts of it. Some forks of roop do this already but they mostly just concentrate on GFPGAN only. I did it to as an experiment with DMDNet, that's why it is so fast. |
Beta Was this translation helpful? Give feedback.
tldr; The current enhancers except DMDNet don't allow parallel execution.
Yes it's possible and there's nothing wrong. The problem is that all enhancers in their default form don't like running in parallel. This means that your face swapping is properly threaded but after that every thread has to wait in order for post-processing. Also every face restoration does the face recoginition process again. I know this is stupid but to improve this you would have to dissect the existing code and rewrite parts of it. Some forks of roop do this already but they mostly just concentrate on GFPGAN only. I did it to as an experiment with DMDNet, that's why it is so fast.