{"version": "1.0", "type": "rich", "title": "the thing about the deepfakes videos\n (some hobbyist trained publicly available machine learning so that given an archive of...", "author_name": "kontextmaschine", "author_url": "https://kontextmaschine.com", "provider_name": "kontextmaschine", "provider_url": "https://kontextmaschine.com", "url": "https://kontextmaschine.com/post/168560433443/", "html": "

kontextmaschine:

\n

the thing about the deepfakes videos

\n

(some hobbyist trained publicly available machine learning so that given an archive of celebrity photos it could paste that face into porn videos, surprised you guys weren\u2019t more interested)

\n

is that a face isn\u2019t really THAT much. I guess people were photoshopping celeb faces into nude photos circa AOL (and also, photoshopping superhoeroine costumes onto them? dunno, man) and that worked

\n

But back then there were porn lookalikes, that was maybe more of a thing, but they\u2019d try. Like, if you really looked like ScarJo (or anywhere close), you\u2019d come done up the Black Widow hair and outfit, not looking like a Florida sorority stripper

\n

he says doing just the face saves cycles on a consumer rig, he does the Taylor Swift and Emma Watson ones on the same substrate and more than the body it\u2019s weird watching them matched to the same hair

\n
\n

and it\u2019s interesting, as far as I can tell he does it by taking a pile of celebrity photos, and then running them through some program that produces distortions of each one, and then training the program to correct the distortions

and after it\u2019s gotten skilled enough at that he can loose the program on the substrate video, and for every frame wherever it sees a face it corrects it back towards the celebrity

"}