Has anyone figured out what this is yet? I still see people saying different things so I don't know if it's still a bit of an experimental "well this way seems to work for me so I'm sticking with it" - without realizing maybe 50% of the faceset doesn't make any positive contribution to training just slows it down being processed. I can't find it now but someone on here said they'd made a deepfake from only 40 or 60 images? but I figure most people still grab several thousand.
So 40-60 is probably too small, but 5000 probably too big - it's almost like you need footage from when they filmed the matrix with the 360 camera rig around the actor - and could say harvest X images from angle X to Y and repeat whilst they did a different facial expression you could get to the critical say 500 images that the algorithms could then get all they need from it, anything beyond that doesn't add value.
I notice in another thread Samsung supposedly did one from 1 image, so we must be at a point we don't need such big facesets, where is the point of diminishing returns if we can define the "essential" base set of angles/expressions.
So 40-60 is probably too small, but 5000 probably too big - it's almost like you need footage from when they filmed the matrix with the 360 camera rig around the actor - and could say harvest X images from angle X to Y and repeat whilst they did a different facial expression you could get to the critical say 500 images that the algorithms could then get all they need from it, anything beyond that doesn't add value.
I notice in another thread Samsung supposedly did one from 1 image, so we must be at a point we don't need such big facesets, where is the point of diminishing returns if we can define the "essential" base set of angles/expressions.