Trained face weirdness

Got questions or tips about the Conversion process? This is the place to discuss them.


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 Converting faces from your trained model.

If you are having issues with the Convert process not working as you would expect, then you should post in the Convert Support forum.

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

Locked
User avatar
rocketpants
Posts: 2
Joined: Sat Oct 24, 2020 8:29 pm

Trained face weirdness

Post by rocketpants »

I'm having some trouble figuring out an issue that's occurring during training and would love some help. Please take a look at the 3 images below for examples of what is happening.

Image

This is after over 100k iterations training using Dfaker. I've also seen this using Realface. I have trained with Dfaker to over 300k iterations on a previous attempt and while the swap looked amazing, I still had this blurring on the edges. What's causing this? Is it the training settings, or is it the alignments? I have tried so many settings changes to see if I can find the cause but so far I'm having no joy.

User avatar
torzdf
Posts: 2651
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 129 times
Been thanked: 622 times

Re: Trained face weirdness

Post by torzdf »

This is the edge of the mask, and can become particularly pronounced when face shapes are significantly different.

However, it can be fixed, to an extent. Use the preview tool to get the blending settings correct. You can see the newly added convert guide to get more info:

viewtopic.php?f=7&t=1083

My word is final

User avatar
rocketpants
Posts: 2
Joined: Sat Oct 24, 2020 8:29 pm

Re: Trained face weirdness

Post by rocketpants »

Oh cool, I didn't know there was a Convert guide now. That's great, thanks. I read through it. I might have been making a couple of mistakes in the Preview and Convert stages so I'll test those out now that I have a clearer understanding of how Convert works.

So if the face of inputA is wider than inputB (as appears to be the case in my example), the type of Extract of those faces, or the type of training model, probably aren't going to do much to fix that, do I have that right? The only thing I can try to do is adjust the blending mode before converting.

User avatar
torzdf
Posts: 2651
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 129 times
Been thanked: 622 times

Re: Trained face weirdness

Post by torzdf »

Pretty much, yes. Face shape can be tough sometimes.

My word is final

Locked