I’m currently away from the computer at the moment, but when I was on it a couple days ago I was having an error with both extractions and training when using older models.
It started off first with me wanting to try and train some more on an older project I’m doing on my spare time between paid requests, I had some new A data I thought would be vital for training better, so I extracted it (same image pixel size) and when I tried to click train it told me that it wouldn’t, there was some type of issue with my B data. It wanted me to re extract my B data. I understand updates break things, it’s a pretty small issue so I tried re-extracting my older B data so it can train and the images all came out really small, despite the same extraction size being selected. And when I mean small I mean the file size was still 256 by 256, but there was a huge black border all around it and the image was unnecessarily small. To give you an example here’s the image of what I mean. On the left is the old B data extracted at 256 size and it’s exactly how you would expect it. While on the right is the brand new B data where the file size claims it’s still 256x256 but it’s really small. (Faces censored because I don’t have the paid requesters permission yet to uncensor it)
And I want to stress there’s nothing wrong with the right picture, it should extract and look just like the pic on the left. It’s not like really small and that’s the best faceswap could do.
I became instantly weary, so I made a copy of my model to try training this new A data I had and the re-extracted but inexplicably small old B-data, because I was worried something would go wrong and I was right. When it started to train it got real
Blurry like I introduced a brand new face into the A-B equation. And this blurriness stayed even after training for like 8 hours.
What can I do to fix this? Is this an error related to the update a couple days ago? Was there just some hidden setting I had on when extracting that caused my pictures to be really small? (Even though I had the correct 256 size on)