Faceswap Forum

So, I'm 1.5M iterations into a training session and I'm not getting the quality I would expect. Good enough, I check the usual culprit of input quality. I found that I somehow missed two totally misaligned images in a batch of 1,300. I've updated the alignments and removed those PNG's from the training sets. I know the former isn't strictly necessary but I'm reusing a source set to better train myself on how to do this successfully. Someday, perfection in and perfection out, or so I hope.

How impactful is having a bad alignment image (or two in this case) in a 1,300 image batch? Does it taint the entire set and I should start over? Is it irrelevant and removing them was a waste of time? Thoughts?

Second, and not really related... How (on windows/PC) do I get the help menu to display during manual alignment? I don't see a control for it. Is there a hotkey?

Thanks for your thoughts...

It's a percentage game. You are generalizing a model. A very small percentage of misaligned images will have a very small percentage impact on the model.

If you're not getting the results you want the issue is, more often than not, not enough variation in the datasets.

ref: viewtopic.php?p=7506#p7506