(first time user) 30k iterations - not much difference yet

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
paranoid_android
Posts: 4
Joined: Sat Feb 20, 2021 9:14 am

(first time user) 30k iterations - not much difference yet

Post by paranoid_android »

Hi guys,

I'm a newbie here, so my first swap is very likely going to be awful. But I just want some guidance on whether I'm going in the right directions or to give up and start from scratch on this first project.

As you can see from the image below, the faces are barely being swapped as yet. There has been over 30k iterations, so that's still very low. But at what point should I begin to see some merging of faces? I thought by now there would at least be a hint of faces being merged, albeit a bit blurry.

I'm using the original trainer, batch size of 48 (I had been using batch size of 12 for most of the training, but decided to pauze and increase the batch size and resume).

Running on a macbook pro - 2.4 GHz 8-Core Intel Core i9, AMD Radeon Pro 5500M 8 GB.

Image

Should I keep going with this for longer?

User avatar
torzdf
Posts: 2649
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 128 times
Been thanked: 623 times

Re: (first time user) 30k iterations - not much difference yet

Post by torzdf »

There is no way near enough variety in your data here. Lighting sources specifically.

app.php/faqpage#f3r0

My word is final

User avatar
mrdeeplearning
Posts: 3
Joined: Sat Feb 20, 2021 11:06 am
Has thanked: 1 time

Re: (first time user) 30k iterations - not much difference yet

Post by mrdeeplearning »

[mention]torzdf[/mention] is correct. Please take a HQ video of yourself under various lighting conditions and angles.
Also, if I am understanding your preview correctly, you want to swap your face with Kanye. So Kanye would be A and you would be face B

The reason you are not seeing ANY difference is b/c you have the order swapped. The model is learning your face and swapping it back onto you. Swap the order and BOOM, you will see it being swapped. Also, which model are you using? Depending on that, you might still be able to swap your outputs.

[mention]torzdf[/mention] correct me if I am wrong, but [mention]paranoid_android[/mention] can probably switch the decoders while converting right?

Locked