Faceswap newbie here. Just started working with the software today and could already get some first impressions as to how it works and what it can do. The main limiting factor at this stage was obviously the fact that within the span of just a few hours I wasn't able to do training sessions that were substantial enough to produce really good results. (Got problems getting it to work with my GPU, too. But that's a topic for another day. My CPU seems to work reasonably fast in the training sessions for now.)
Earlier today, I shot some material of my own face with my camera on the tripod to work with. I intend to do the same with some friends in the future. This automatically raises the following question: If I'm not restricted to using whatever suboptimal source material I can find of e.g. a public figure, but if I get to shoot my own material in a controlled environment (at least for one part of the two-part equation), what material am I going for with regard to the poses/perspectives that I'm getting of the face. I mean, I suppose it also depends on the other source video that I want to replace the face in, and whether that features any "extreme" perspectives that differ a lot from just frontal/portrait footage. However, if the intention is to just have a fairly generally useful base to build on that I can use as swap material for a lot of videos with frontal face footage, what would you recommend? Do you have a preferred "posing routine" for the video? Based on other recommendation regarding the number of pictures you want to work with and have after the extraction, I figured I should go for something between 30 and 60 seconds (shot at 24p). Naturally, I try to have all the other, more technical aspects dialed in that I would pay attention to for any controlled shoot: focus, lighting, reasonably good image quality and bitrate. However, what should one be doing in front of the camera to have good source material? Obviously you don't want to be too static. I suppose you'd want to get footage of your face from slightly different perspectives, tilting the head up and down just a little, turning it left or right a few degrees? (How far?) What else? Talking? Moving your eyes?
So, in short: Is there anything as a prescribed "posing routine" that would give the algorithm good material to work with?