Hmm, I thought deep fakes consume lots of computational power both at training and inference time. Is there some tech/science breakthrough here or they have just scaled it up to more hardware?
So I'd expect it to only work on the 1 person in the demo, with the 4 famous people depicted.