Faceswap Forum

I wasn't sure how to phrase this in the title, but I'm wondering how good of a strategy would be to start training with no so good sources the first 20k iterations, and stick to the best ones for the remainder of the process.

Right now I have 1400 very good faces of B, with a wide variety of lightning conditions, but they're lacking some angles. I'm thinking of adding some 600 not so great faces with those missing angles, and drop them as I hit the 20k iterations. My intention is to provide the main forms of the face at those angles, but I don't know how much it would hurt the model.

What do you think? I want to try it out unless you tell me it's decisively a bad idea with a certainty of failure.

This is an aspect of what I call "fit training". See the guide viewtopic.php?f=27&t=74 for some more explanation of the process. Basically once fully trained, you train on just the swap video (and in your case, your best data for B) and let it try to master the detail.

Great, thanks!

I actually read that guide (I've been trying to read everything on the forums, even the technical stuff I don't get), but I didn't realize this approach could be considered a particular case of fit training.