This should use a still image perceptual comparison algorithm to compare output. Something similar to PSNR.