I was proposing to skip the (computationally expensive) step of approximation of the illumination effects on the subject within the live deepfake model, and jus

I was proposing to skip the (computationally expensive) step of approximation of the illumination effects on the subject within the live deepfake model, and just do a (computationally cheap) filter over the webcam image to be sent out. I was saying 'correctly faking' might not be necessary, so let's fake it incorrectly. Of course I haven't tested it, but from reading the article, this seems to be enough to fool their detection model.