An apparatus and method for predicting human vision perception and perceptual differences provides a perceptual difference model that merges two human vision systems, one for a reference video signal and the other for an impaired version of the reference video signal. The respective video signals are processed by spatio-temporal filters and, prior to differencing, by noise masking modules that apply a noise mask as a function of pupil size and luminance. The differenced filtered signal is an initial perceptual difference map to which other masking is applied to take into account correlation and contrast gain based upon the noise masks and filtered luminance from the spatio-temporal filters. The result is a more precise output perceptual difference map.