that's quite cool, though when applied to irl with complex lighting it's not doing right
to be honest, i believe that creating a filter that would work exactly like that (so follow the edge of body then fill it with whatever) would require an another approach, i wonder if some kind of vector output is possible from this kind of neuralnet models - if the detection model would be able to return a vector shape of body, then having this kind of clipping effect should be very easy. I wonder though how much more resource-hungry this model would be, I guess that having a vector image processing would also require a huge change in the whole program's code and the way that image is processed now - but those are only guesses 'cause i've got little to no knowledge how those things work