trained a lot of models and images i got some practice suggestion

Want to understand the training process better? Got tips for which model to use and when? This is the place for you


Forum rules

Read the FAQs and search the forum before posting a new topic.

This forum is for discussing tips and understanding the process involved with Training a Faceswap model.

If you have found a bug are having issues with the Training process not working, then you should post in the Training Support forum.

Please mark any answers that fixed your problems so others can find the solutions.

Locked
User avatar
cuplearner
Posts: 6
Joined: Wed Jan 20, 2021 9:29 am
Been thanked: 1 time

trained a lot of models and images i got some practice suggestion

Post by cuplearner »

to get a comparable good result i followed the following rules
1.the most important try to get the faces either original or swap that have the same shape of face.
2.the expression of face set A that you want to swap with must exists in the face set B that you want to swap on.this can guarantee the decoder to konw how to swap the face set A by learning the most similar expression on face set B.
3.train at least 30k you will get a good result.

above,in summary is data quality.

User avatar
cosmico
Posts: 95
Joined: Sat Jan 18, 2020 6:32 pm
Has thanked: 13 times
Been thanked: 35 times

Re: trained a lot of models and images i got some practice suggestion

Post by cosmico »

Those are some pretty good observations that I've noticed myself and I've been playing around with faceswap for over a year. I can add on to what you are saying with some of my own observations.

Assuming your data on both ends are videos and there moderately high quality like 720p or above, open the videos in a video editing program like shotcut, premier, vegas and add some sharpness to the video before you extract it. It helps simulate higher quality data when you do that. Make sure you leave some normal and blurry data though so it know hows to deal with blurs.

Also if you want your data to train even faster, then make it so that your B data doesn't teach the model anything it doesn't need to know. For example, If A video's model never opens there mouth, You can speed up the training process by not bothering allowing the model to learn how to open B data's mouth. Every single second the model spent learning how to open B's mouth when it wasn't needed could have been a second it was learning more important things instead. Of course this is a double edge sword, the second you introduce an A video where the mouth opens, the model will be significantly lacking in the understand on how to do that. Also its easy to make mistake and not realize you didn't provide sufficient enough data for it to learn.

User avatar
cuplearner
Posts: 6
Joined: Wed Jan 20, 2021 9:29 am
Been thanked: 1 time

Re: trained a lot of models and images i got some practice suggestion

Post by cuplearner »

8-) yeah,I really agree with you.In fact i am not very familiar with the models、loss and something whatever。I just try to understand what the models do.I think the models just learn to recreate the faces that they have seen,within the shape of face. the eyes、nose 、mouth and expression are the materials. The swap actually swap the eyes and nose then remain the expression.

Locked