Hi everyone, I have finished my first deepfake test. The result as you can see is poor. I ask you if you can help me to understand where I went wrong. It makes me think that the original video file (1280x780) was not of sufficient quality.
I used these settings:
size extract 256
train BS64 -140,000 iterations
model original
As you can see from the last timelapse it was already evident that face B was blurred compared to A and my chin remains out of the red grid. Is it a problem? How to remedy?
Maybe I need to do more iterations? I am also attaching the graph, after 140,000 iterations:
Original video:
Me:
Final result:
update
Ater 250,000 iterations, not Improvements.
Re: My first, poor, deepfake
Posted: Wed Sep 23, 2020 11:48 pm
by FuorissimoX
Someone can help me?
Re: My first, poor, deepfake. What i wrong?
Posted: Thu Sep 24, 2020 9:10 am
by torzdf
Ultimately, it comes down to data. You need lots of varied data for both sides, not just from a single source for each.
Re: My first, poor, deepfake
Posted: Thu Sep 24, 2020 10:07 am
by FuorissimoX
I have about 3,000 pics for side.
Maybe are too similar? You mean I need different pics for each?
Its better if i use 1/4 NN frame (1/1 are too similar) but from different source?
I have see this tutorial and i do exatly the same but my results its bad.
After 50h (using P100) im really frustrated
Re: My first, poor, deepfake
Posted: Thu Sep 24, 2020 10:33 am
by torzdf
Yeah, that guide is a good starting point, but it only covers the basics.
You want lots of varied data from lots of different sources.