Faceswap Forum

Hi all,

I am wondering if there are drawbacks to adding additional training data part way through training.

Suppose after many iterations and despite low loss values your preview still looks blurry or bad, but the data you've used is of decent quality, though perhaps insufficient. If you could add new data, is there any potential for a worse outcome?

Thanks!

Yes, go ahead and add data. That's fine and a normal part of training.

bryanlyon wrote: ↑Thu Nov 12, 2020 7:11 pm

Yes, go ahead and add data. That's fine and a normal part of training.

Thanks - could you comment on the best workflow for this? The first time I did it I didn't know what to do with multiple alignment files, and I ended up re-extracting from a folder of extracted images which created its own issues.

Suppose I found new, high-quality footage - what's the best approach to adding to my existing data set? Apologies if this has been discussed elsewhere - I searched but didn't find anything about it.

Thanks!

You should merge the alignments files when adding new faces.

see here:
app.php/faqpage#f3r4