Do trainer options (e.g. flip chance) apply only to "Original" model or to everything?

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
Replicon
Posts: 50
Joined: Mon Mar 22, 2021 4:24 pm
Been thanked: 2 times

Do trainer options (e.g. flip chance) apply only to "Original" model or to everything?

Post by Replicon »

In the config settings, there are global options (applies to everything globally), model options (applies to specific models), and trainer options.

Do the trainer options apply regardless of model used (meaning, the "Original" trainer has nothing to do with the "Original" model other than the fact that they were the first ones released), or do those options apply selectively? I assume it applies to all models, but I thought I'd ask, to determine if it's worth manually adding some horizontally-flipped sections when I pre-edit my source videos... like, if the subject is mostly front-facing, but glances to the left a few times, I might want to duplicate the "glance" clips with a h-flip if the trainer doesn't already do that in all cases. I guess it might be worth doing anyway, if I want the trainer to have a slightly higher chance of selecting a "side-glance" frame, to better round out the training data.

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

Re: Do trainer options (e.g. flip chance) apply only to "Original" model or to everything?

Post by torzdf »

It does it for all models.

It's structured the way it is as the "original trainer" is a plugin, and it leaves us scope to add other trainers if we ever see a need.

So, yeah, no need to manually process any faces. Also, manually processing is likely to strip the alignment data from the image header, so you probably don't want to be doing that anyway.

My word is final

User avatar
Replicon
Posts: 50
Joined: Mon Mar 22, 2021 4:24 pm
Been thanked: 2 times

Re: Do trainer options (e.g. flip chance) apply only to "Original" model or to everything?

Post by Replicon »

Sweet thanks for confirming!

Oh yeah I wasn't talking about manually tweaking the extracted photos, but literally appending some extra flipped clips at the end of the source video (with Kdenlive, say) and then using that as the source for the initial s3fd extract. Since the trainer already does the flipping, doing flipping in the source video is definitely not useful, but duplicating/lengthening/slowing "rare" clips might still add value (e.g. if the person only blinks once, pasting/lengthening the blink might get more "blink" frames selected during training iterations).

Locked