Could be from simple upscaling techniques (e.g. bilinear interpolation) combined with antialiasing. CNNs don't seem feasible given the compute power needed for the number of users using YouTube.