AliaksandrSiarohin

You video source image is not cropped, try to crop it first.

No it will not help.

Maybe you can try this out, but not sure that it will help. I did not check. https://github.com/AliaksandrSiarohin/first-order-model/issues/146#issue-624354694

I tried reduced training (20 epochs) and it seems to work.

It will take significant amount of time, for example generation all videos for vox test set is 8 hours. Also training loss has pretty good correlation with final accuracy, so...

Sorry, but your code makes no sense to me. You can found how validation is implemented in reconstruction.py

I don't understand what is user defined pose, how it is resented and how the user should get one?

Why not to use the same architecture than?

Yes, so why can't you use their same architecture? In one of their ablation studies they do like this.

Not familiar with it, maybe someone else could help you.