A full-reference metric compares the impaired video against the complete pristine original, frame for frame and pixel for pixel, the way a proofreader checks a copy against the source document with both pages side by side. PSNR, SSIM, MS-SSIM, and VMAF are all full-reference, which is why they are the most accurate family and generally correlate best with human ratings, and why they own the lab: encoder comparison, regression suites, CI/CD quality gates, and video-on-demand QC. The catch is twofold. First, you must possess the master, so full-reference is useless on a live stream or user-generated clip that has no original. Second, the two videos must be aligned in time, space, and brightness; a small temporal, spatial, or luminance offset makes the metric report a quality collapse the eye would never see, so an unaligned score is broken, not low. Full-reference is the richest of the three reference setups, with reduced-reference and no-reference giving up information for reach.
