Faceswap Forum

First time faceswapper so be kind. Using photos as data sets. 1 set worked. Issue is with the 2nd extraction.

Extracted the faces successfully. Sorted them successfully (sorted by 'face', output to 'rename', 'keep' not checked, group by 'hist'). All the faces were sorted.

Then I went into the sorted faces in the output folder and, with an 3rd party image viewer, deleted the bad faces/non-faces. <-- this is the correct way, right?

After deleting the faces in the output folder, i went to files --> tools --> alignments
on 'Jobs' selected 'remove faces' with output as 'console'. Selected the alignment folder in the raw images file folder for the alignment file. Selected the output folder where I deleted the bad face images for the 'faces folder'.

Then I clicked 'alignments' to run it and the process ran and completed.

All this seemed to work like the first round of photos which did work.

However, when I launched the manual tool for fixing the alignments, it immediately became obvious that the wireframes for the faces were for completely different images. Eg. An image of with a small face looking left has a wire frame for a full frontal closeup image. All of the frames and images were completely mismatched. It seems obvious that the alignments file is out of sync with the image files.

So my questions are: 1) how do i correct this? 2) how do i avoid unsnycing the images from the alignment file?

Are you able to provide a screengrab illustrating your issue (a short screen capture video would be even better)? It's been a long time since I used images as a source, but that really shouldn't matter

I have noticed a similar issue. When I extract and then move to manual align, sometimes I notice clicking on a face that looks aligned and OK suddenly shifts the mask and alignment to a different location within the frame. This seems to happen more often when I use the same session to extract from multiple sources right after each other. I now shutdown the application and relaunch between extracts and it has been my workaround for this issue. Not sure if it is the same for your case.

Here is some typical examples. The alignments are obviously not for the same image at all, for all of them.

Hmmm, ok. That is odd. Do the faces previews look correct?

Ultimately, I think I'm going to need you to zip up your pics and alignments file and provide for me so I can see what's going on here. You are welcome to PM me the link.

Nope. The previews in the thumbnails are representive of the very off wire frames :/ I redid the extraction too, did it again and had the same thing happen

Ok, the only way I'll be able to investigate this is to examine your source data and alignments file. If you can zip them up and provide to me (by PM) then I can investigate. I will only be using this data to fix this bug, I can guarantee it will not be shared.

I think I figured it out. There seems to be a bit of a glitch I tested a few times.

Since I was using photos for the sets, in the extraction process I had set N=2 because I wanted half the set done so I could test the results on the other half or use those images for training. However, this seems to throw the alignments file completely out of wack. The fix is simply set n=1 when extracting.