Training Villain

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
snappy
Posts: 8
Joined: Sun Apr 26, 2020 3:54 am
Has thanked: 4 times

Training Villain

Post by snappy »

Hello. :D

I hit 412k iterations on Villain @ batch size 26 with average 21.6 EGs/s

I do not think that it still learns at this point and the results aren't the greatest. Not close to the quality I see on some youtube examples. I will leave it go for another 100k iterations.

I was wondering if there's a way for any of you experts to take a look at the faces/settings/sources I used for this swap. I feel like I'm real close to getting that great, clean swap but I need some fine tuning/help.

Image

User avatar
djandg
Posts: 43
Joined: Mon Dec 09, 2019 7:00 pm
Has thanked: 4 times
Been thanked: 2 times

Re: Training Villain

Post by djandg »

The more you train the less the return, especially once over a few thousand It's x EG;s.
It's not linear, so training for another 25% of time doesn't gve you a 25% better swap. As you summized, learning reduces the more you train.
I've done experiments at 450k, 650k and 1 million and the gain in swap quality between 650k and 1 million is small, best balance was at 650k. At 1 million it had still not converged but training more was a waste if time.
The cure to get better is many many more face images of sharp quality from sources at the right size to match what you want to swap (close up, mid field or far away) to feed in to the training.

Image

User avatar
snappy
Posts: 8
Joined: Sun Apr 26, 2020 3:54 am
Has thanked: 4 times

Re: Training Villain

Post by snappy »

Yea, I feel there's more than selecting the perfect cuts. This guy knows something we don't.

I'm putting emphasis on the shine in the eyes. Brilliant quality.

Last edited by snappy on Thu May 14, 2020 11:01 pm, edited 1 time in total.
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: Training Villain

Post by bryanlyon »

Lots of post processing in external software.

User avatar
snappy
Posts: 8
Joined: Sun Apr 26, 2020 3:54 am
Has thanked: 4 times

Re: Training Villain

Post by snappy »

Can you give me some examples of post processing tools ?

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: Training Villain

Post by bryanlyon »

He lists one himself in the description of that video. Hitfilm Express. But it's not going to be a "drag and drop" or one click solution. It's something that you'll actually have to get experience, practice and effort in.

User avatar
snappy
Posts: 8
Joined: Sun Apr 26, 2020 3:54 am
Has thanked: 4 times

Re: Training Villain

Post by snappy »

I'm glad it's not drag and drop, if that was the case everyone could do it ;)

Locked