Some folks at NVIDIA have developed a "style-based generator architecture for generative adversarial networks", borrowing from "style transfer literature". It has been demonstrated to generate remarkably realistic (and fictitious) human faces. As impressive as it is, it's not obvious the extent to which the demonstration of this architecture is like a little old lady who only drove it to church on Sundays. Regardless, it looks like the AI overlords can now have alleged supermodel girlfriends in Canada too:
The video demonstration is at https://www.youtube.com/
Their explanation/description is at https://arxiv.org/pdf/1812.04948.pdf
(Score: 2) by crafoo on Friday December 14 2018, @05:55PM (2 children)
The future sure is rushing up on us pretty fast these days. I suppose this could be applied to a 3D modeled mesh of various faces (and bodies) as well, but the training sets would have to be fully 3D modeled.
(Score: 0) by Anonymous Coward on Friday December 14 2018, @07:03PM
The SOTA is pretty good for portrait to 3D model (including hair), so these images could just be plugged into that.
(Score: 2) by takyon on Friday December 14 2018, @08:36PM
Check out Two Minute Papers [youtube.com]. Then keep checking it out every week.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]