Faceswap Forum

Hello,

I've spent the last 2 weeks extracting, aligning and masking clips for a face swap. I completed all of face B (the face to be applied) a few days ago and face A yesterday. I had originally intended to merge the alignment files for each face and train against all of the faces but I see the new way has that information inside the png files themselves.

I set about the process of recreating the image sets from the original alignment files for every video source I used (280 sources, took some time) and tried to start training but I received this error:

ERROR    Mixing legacy and full head extracted facesets is not supported. The following side(s) contain a mix of extracted face types: ['A']

The error didn't provide any more info about which particular source images might be causing the problem so I began the task of trying limit sets at a time (also took a long time). At a certain point I finally clued in. All of the "new" extractions I did can't be used with the old extractions... Ah, that's... that's not great. If I can't use the "legacy" facesets for face A I will have lost days of wok. Let alone all of face "B" will be legacy.

Is there any way to upgrade legacy extractions? Can I train with a legacy and then a non-legacy session? Can I use any of the 13.500 extractions for face B?

This may have been a bug in the legacy/full-head extract test. I have pushed an update to fix.

Please let me know if it fixes this particular issue.

Otherwise it just is not possible to support training on 2 different types of extract on the same side. The amount of code required to support a narrow and diminishing use case would not be worth it. If you still have the issue, I can maybe look at ways of getting a list of legacy extracted faces out. However, it should be apparent from visually scanning the folder, as the "zoomed in" faces will be legacy extracted.

The alignment tool can be used to re-extract faces to the newer method.

Thanks for your reply. I used the alignment tool to re-extract all the images before my post it didn't work then but I'll try again if the test still fails.

ETA: your bugfix worked. I am able to use the re-extracted images. Thanks