Read the FAQs and search the forum before posting a new topic.
This forum is for reporting errors with the Training process. If you want to get tips, or better understand the Training process, then you should look in the Training Discussion forum.
Please mark any answers that fixed your problems so others can find the solutions.
This pic is from a transition where the head is tilted pretty far back. On the right, you can see the distended blob that it became which is what always happens when I do faceswaps from any source video. I doublechecked my face b pics and alignments and they seem fine. The alignments on the source seem fine. I don't know what's going on. I've tried Realface (600000), Villain (400000), and Dlight (250000), but no luck.
I'm also having a weird problem with one eye being open and the other closed...
At several points during this video (haven't seen this issue before - or wasn't too noticable), the eyes are looking in the wrong direction or one eye is weirdly closed when it shouldn't be.
This occurs because you don't have enough/any data in B representing this angle, so the model cannot swap it effectively. If you look, A can recreate this fine, B cannot. That is due to insufficient data covering that pose.
2 options:
re: eyes. These can be tricky, Some models handle it better than others. I know that dlight + realface can have issues with eyes. It may be down to the unbalanced nature of the models, it may not be.
I don't have such a direct solve for this one. Generally ensuring data covering eye directions wouldn't hurt. However why a model may learn that eye direction is one way for the A side and different for the B side I couldn't say due to the 'black box' nature of Machine Learning
My word is final
Fair enough. So I need more source material with that angle. You're right that it's model-dependant on the eye thing though. I've only ever seen that dlite.
