Better A > B Swap than B > A on different trainer models (Newbie)

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
MetaUserName
Posts: 6
Joined: Mon Mar 09, 2020 7:48 pm
Has thanked: 1 time

Better A > B Swap than B > A on different trainer models (Newbie)

Post by MetaUserName »

I've recently begun working with the Faceswap program and have had trouble getting good results when swapping face B onto A, while getting great results swapping A onto B. I have tried Dlight, Dfl-h128, original, and lightweight and am getting the same results across each. In general these results are:

  1. Face B swapped onto face A has poorer results than Face A swapped onto B.
  2. The loss for A is always greater than the loss of B.

The computer i am using is farily low-end (Nvidia Georce GTX 1050) so i am using both memory saving gradients, optimizer savings, as well as running batches no larger than 30.

Is their some trick i am missing during the conversion/training or is it common for A to have greater loss than B?

User avatar
bryanlyon
Site Admin
Posts: 793
Joined: Fri Jul 12, 2019 12:49 am
Answers: 44
Location: San Francisco
Has thanked: 4 times
Been thanked: 218 times
Contact:

Re: Better A > B Swap than B > A on different trainer models (Newbie)

Post by bryanlyon »

For Dlight, this is an intended "feature" of the model. It saves memory on the A decoder in order to maximize the quality of the B side. This also applies to Realface and Unbalanced.

All the other models you mentioned are symmetrical where the A and B sides are identical and you can swap both directions. If you're seeing higher A loss on these models, it's probably due to your data.

User avatar
MetaUserName
Posts: 6
Joined: Mon Mar 09, 2020 7:48 pm
Has thanked: 1 time

Re: Better A > B Swap than B > A on different trainer models (Newbie)

Post by MetaUserName »

bryanlyon wrote: Tue Mar 10, 2020 4:33 pm

For Dlight, this is an intended "feature" of the model. It saves memory on the A decoder in order to maximize the quality of the B side.

I appreciate you mentioning this, the description for Dlight does not seem to indicate that it would do this.

I have continued to run my project and i am slowly getting better results using the dfl-h128 model. Face A still has a higher loss than face B, but face B has better lighting.

User avatar
Grassone
Posts: 45
Joined: Sun Apr 19, 2020 7:32 pm
Has thanked: 6 times
Been thanked: 2 times

Re: Better A > B Swap than B > A on different trainer models (Newbie)

Post by Grassone »

Hmmm wait...wait...wait.... this is kind of confusing..

The trainer window looks like this (at least this is what I see on the pop-up balloons)

Does this mean that I am doing things the wrong way (If I want to keep a good quality)?

Attachments
Hmmm.JPG
Hmmm.JPG (24.64 KiB) Viewed 4962 times
User avatar
torzdf
Posts: 2649
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 128 times
Been thanked: 623 times

Re: Better A > B Swap than B > A on different trainer models (Newbie)

Post by torzdf »

Your annotation is correct.

My word is final

User avatar
Grassone
Posts: 45
Joined: Sun Apr 19, 2020 7:32 pm
Has thanked: 6 times
Been thanked: 2 times

Re: Better A > B Swap than B > A on different trainer models (Newbie)

Post by Grassone »

Sheeesssss...

This explains a lot of things... including -probably- the fact that randomly there are faces on the "weak" side looking like Marty Feldman...

Anyway, this had a good side effect: now I am cleaning my input data a lot more carefully.

Locked