You are viewing a single comment's thread.

view the rest of the comments →

0
4

[–] FeepDakeNoob 0 points 4 points (+4|-0) ago 

I think the three most important things are:

1) quality of training data: Blurred or obstructed faces should be tossed from your training set. A few hundred quality images will get you better results than 5,000 crappy images.

2) model selection: You touched on this one and it's important. Your models should be as similar as possible if you are aiming for realism. Head shape is probs the most important aspect of this. Trying to swap a face from a round faced person to someone with a long narrow face is not ever going to look good. Skin tone can be a factor too.

3) Realistic expectations: Thinking that this software is perfect and ready to go out of the box is a total fantasy. We're a long way off. Thinking that you can dive into this without some intermediate knowledge and computer troubleshooting skills probably will be a very frustrating journey for you. Also, thinking that you can crank out a perfect video in a day using a graphics card from 2014 is a mistake. This stuff takes a lot of time and trial and error.

0
0

[–] Robjones1234 [S] ago 

Good points all around! I also think people don't try to vary the celeb facial expressions enough. The worst is when you got a good deepfake going and she leaned back and the effect is totally lost. I suppose if the pornstar looks similar already, it may mask this but when it's bad... ITS BAD.

Anyway, rdcy fakes on Tumblr found a scene where Gal Gadot tilts her head back and closes her eyes. He used it for a headswap but it's exactly the kind of thing more deepfakers need to have in their arsenal if they are ever going to take it to the next level