Functional connectivity analysis has become a powerful tool for probing the human brain function and its breakdown in neuropsychiatry disorders. So far, most studies adopted resting state paradigm to examine functional connectivity networks in the brain, thanks to its low demand and high tolerance that are essential for clinical studies. However, the test-retest reliability of resting state connectivity measures is moderate, potentially due to its low behavioral constraint. On the other hand, naturalistic neuroimaging paradigms, an emerging approach for cognitive neuroscience with high ecological validity, could potentially improve the reliability of functional connectivity measures. To test this hypothesis, we characterized the test-retest reliability of functional connectivity measures during a natural viewing condition, and benchmarked it against resting state connectivity measures acquired within the same functional magnetic resonance imaging (fMRI) session. We found that the reliability of connectivity and graph theoretical measures of brain networks is significantly improved during natural viewing conditions over resting state conditions, with an average increase of almost 50% across various connectivity measures. Not only sensory networks for audio-visual processing become more reliable, higher order brain networks, such as default mode and attention networks, also appear to show higher reliability during natural viewing. Our results support the use of natural viewing paradigms in estimating functional connectivity of brain networks, and have important implications for clinical application of fMRI.